Management of Multimedia on the Internet 3540442715, 9783540442714

The 5th IFIP/IEEE International Conference on Management of Multimedia NetworksandServices(MMNS) washeld inSanta Barbara

116 90

English Pages 367 [369] Year 2002

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Management of Multimedia on the Internet
Preface
MMNS 2002 Organizing Committee
Program Committee
Additional Reviewers
Table of Contents
A QoS Network Management System for Robustand Reliable Multimedia Services
1 Introduction
2
QoS Algorithms
3
System Architecture
3.1
Link Emulator
3.2
Metric Collector
3.3
Q-OSPF Daemon
3.4
MPLS
3.5
Applications
4
Experiments
5
Conclusion
References
Federated Accounting Management System Architecture for Multimedia Service Usage Management
1 Introduction
1.1 Related Works
2 Federated Accounting Management Concept
3 Business Context and Functional Requirements Analysis
3.1 Operations Processes Involved in Federated Accounting Management
3.2 Functional Requirements and Use Case Model
4 Federated Accounting Information Model
4.1 Enhancements to IPDR Master Schema
4.2 Federated Accounting Information Service-Specific Using UML Notations
5 System Architecture
5.1 Interfaces (Boundary Objects)
5.2 Components (Control Objects)
5.3 Information Objects (E-IPDR Documents)
6 System Integration and Technology Architecture
7 Conclusions and Future Work
References
Policy-Based Quality of Service and Security Management for Multimedia Services on IP Networks in the RTIPA Project
1 Introduction
2 An Architecture for Policy-Based QoS and Security Management
3 The Policy Refinement Process
4 Network Level Prototyping
5 Conclusion: Open Issues and Perspectives
Acknowledgements
References
Formal Modeling of Service Session Management
1 Introduction
2 A Service Architecture
2.1 Service Session Life Cycle
3 A Service Scenario
3.1 Content Retrieval Service
3.2 Content Retrieval Service Management
4 Modeling Method and Tools
4.1 Modeling Methodology
4.2 Languages and Tools
5 Modeling of Service Management
5.1 Modeling of the Structure of the Content Retrieval Service
5.2 Specifying the Behavior of the Service Components
5.3 Example of End-to-End Service Provisioning
6 Conclusions and Future Work
Reference
Network Requirement for Management of Multimedia over Wireless Channel
1
Introduction
2
System Model and Proposed Retransmission Scheme
3
System Analysis
3.1
Notation
3.2
Average System Failure and Recover Time
3.3
System Model and Dynamics
4
Minimum Receiver Buffer Requirement
5
Numerical Results
6
Conclusions
References
Agile Systems Manager for Enterprise Wireless Networks
1 Introduction
2 Architecture of Agile Systems Manager
2.1 Worker Agent
2.2 Discovery Agent
2.3 Inference Agent
2.4 Management Protocol Component
3 Infrastructure Substrate in ASM
3.1 Pod
3.2 Resource Monitoring Component
3.3 Rule Engine
3.4 Communication Mechanism
3.5 Lookup and Leasing Mechanism
4 Network Performance Management
5 Managed Element Discovery in ASM
6 Performance Monitoring in ASM
6.1 Distance Metric Resource Monitoring Component
7 Observations of ASM
7.1 Fault Tolerance and Load Balancing
7.2 Experimental Observations of Agent Operations
7.3 Security
8 Related Work
9 Conclusion
Reference
System Performance of HiperLAN/2
1 Introduction
2 High Speed Transmission
3 Physical Layer
4 Transport Channels
5 Convergence with other Systems
6 Radio Network Functions
6.1 Dynamic Frequency Selection
6.2 Power Control
6.3 Link Adaptation
6.4 Multi-beam Antenna
6.5 Handover
7 System Performance
8 Comparison
9 Conclusion
References
Streaming Media Congestion Control Using Bandwidth Estimation
1 Introduction
2 Related Work
3 Streaming Media Congestion Control Protocol (SMCC)
3.1 SMCC Sender Operation
3.2 SMCC Receiver Operation
3.3 Bandwidth Calculation in SMCC
4 Experimental Results
4.1 Accuracy of Bandwidth Estimation
4.2 Throughput Behavior
4.3 Fairness
4.4 TCP Friendliness
4.5 Comparison of SMCC to Other Streaming Protocols
5 Conclusion
References
Signaling Protocol for Session-Aware Popularity-Based Resource Allocation
1
Introduction
2
Related Work
3
SAPRA Protocol
3.1
Overview of Protocol Operation
3.2
Example of Protocol Operation
4
Protocol Evaluation
4.1
Theoretical Analysis
4.2
Simulations
5
Conclusions and Future Work
References
A Scalable Video Server Using Intelligent Network Attached Storage
1 Introduction
2 System Architecture
2.1 Data Organization
2.2 Admission Control and Stream Scheduling
3 Design Optimization
3.1 Simulation Assumptions, Parameters and Performance Metrics
3.2 Simulation Results and Analysis
4 Implementation Issues
5 Related Work
6 Conclusion and Future Work
References
On Proxy-Caching Mechanisms for Cooperative Video Streaming in Heterogeneous Environments
1
Introduction
2
System Model and Related Issues
2.1
A Video-Streaming System with Proxies
2.2
Related Issues and Relevant Previous Works
3
A Proxy Cache with Video-Quality Adjustment
3.1
The Cache Table and Remote Table
3.2
Block Retrieval Algorithm
3.3
Block Prefetching Algorithm
3.4
Cache Replacement Algorithm
4
Evaluation
4.1 Infinite Buffer Case
4.2
Limited Buffer Case
5
Conclusion
References
Using CORBA's Advanced Services to Enhance the Integrity of QoS Management Programmable Networks
1 Introduction
2 QoS Architectures
3 Programmable Network Architectures
4 CORBA as a Standard Middleware Solution
5 Management of Correlated NEs
6 Using CORBA’s Transaction Service to Enhance Netw
7 Case Study
8 Using the CORBA Concurrency Service to Enhance Network Management in Distributed Environments
9 Discussion
10 Conclusion
Acknowledgements
References
Remote Multicast Monitoring Using the RTP MIB
1
Introduction
2
Background
2.1
SNMP
2.2
RTP and Related MIBS
3
RTP Sender MIB
4
RTP-Based Monitoring System
4.1
Architecture
4.2
Implementation
4.3
Experience
5
Related Work
6
Conclusions
References
Active Technology as an Efficient Approach to Control DiffServ Networks: The DACA Architecture
1
Introduction
2
Underlying Technologies
2.1 QoS Management Using DiffServ
2.2
Policy-Based Management
2.3
Active Networking
3
Objective of the Work
4
The DACA Architecture
4.1
Lifecycle of Differentiated Services
4.2
Component Description
4.3
Filtering Capsules in the Data Plan
4.4
Monitoring ADR Resources
4.5
QoS Policy Definition
4.6
TCB Definition and Deployment
4.7
Sensor Deployment
5
Prototype Implementation
6
Conclusion and Future Work
7
Acknowledgments
8
Bibliography
Efficient Implementation of Packet Scheduling Algorithm on High-Speed Programmable Network Processors
1
Introduction
2
Related Work
3
DWCS Algorithm and Earlier Implementations
4
IXP Architecture
4.1
IXP Hardware
4.2
Pipelined Processing
5
Hierarchically Indexed Linear Queue \(HILQ\) –
5.1
Drawbacks of the Heap Structure
5.2
Assumptions
5.3
Data Structures
5.4 Pipeline Stages
5.4.1
Receiving Threads
5.4.2
Scheduling Thread
5.4.3
Garbage Collection Thread
6
Performance Evaluation and Analysis
7 Conclusion
8
References
Distributed Dynamic Capacity Contracting: A Congestion Pricing Framework for Diff-Serv
1
Introduction
2
Related Work
3
Distributed-DCC: The Framework
3.1
How to Calculate $p_ij$ ?
3.2
Budget Estimation at Ingresses
3.3
Congestion-Based Capacity Estimation at Egresses
3.4
Capacity Allocation to Edge-to-Edge Flows
3.5
Fairness
3.6
Scalability
4
Edge-to-Edge Pricing Scheme (EEP)
5
Simulation Experiments and Results
5.1
Experiment on Single-bottleneck Topology
5.2
Experiments on Multi-bottleneck Topology
6
Summary
References
Fair Stateless Aggregate Traffic Marking Using Active Queue Management Techniques
1
Introduction
2
Related Work
3
Traffic Marking
4
Probabilistic Aggregate Marker
5
Fair Stateless Aggregate Marker
6
Experimental Results
6.1
Experimental Setup
6.2
Validating the Marking Process
6.3
End-to-End Performance
7
Conclusion and Future Work
References
A Dynamic Marking Scheme of Assured Service for Alleviating Unfairness among Service Level Agreements
1 Introduction
2 Two Previous Works and Rate Adaptive Marching Scheme
2.1 Previous Approaches
2.2 The Axiom of Rate Adaptive Marking
3 Performance Study
3.1 Evaluation of Inter-SLA Unfairness
3.2 Evaluation of Intra-SLA Unfairness
4 Conclusion
References
Minimizing Transmission Costs through Adaptive Marking in Differentiated Services Networks
1
Introduction
2
Adaptive Marking in DiffServ Networks
3
Reinforcement Learning
4
Reinforcement Learning-Based Adaptive Marking (RLAM)
4.1
Motivation
4.2
Overview of Approach
4.3
Measurements
4.4
States and Actions
4.5
Aggregation of Reinforcement and Q-Value Feedback
5
ns2 Implementation
5.1
Network Topology
5.2
Traffic Characteristics
6
Simulation Results
7
Conclusion
References
Dynamic QoS Adaptation Using COPS and Network Monitoring Feedback
1 Introduction
2 Dynamic QoS Adaptation
2.1 Static Policy Decision
2.2 Dynamic Policy Decision
2.3 Automatic Diffserv Domain Configuration
2.4 Example of Application
2.5 QoS Management Algorithm
2.6 Calculating Bandwidth Usage in the Network Links
3 Implementation
3.1 Tool Manager
3.2 Network Monitoring Agent
3.3 Policy Management System
4 Performance Evaluation and Results Analysis
4.1 Experimental Testbed
4.2 Performance Analysis
5 Conclusion
References
Design and Implementation of an Application Layer Protocol for Reducing UDP Traffic Based on User Hints and Policies
1
Introduction
2
Consideration in Design
3
Architecture
3.1
End-User Interaction Monitors
3.2
Client Coordinator
3.3
Server Coordinator
3.4
Interactions
4
Implementation
4.1
Client and Server Coordinators, Policy Table and Interaction Monitors
4.2
Applications and Policies Used in Experimentation
4.3
Testing and Experiments
5
Related Work
6
Discussion
7
Conclusions and Future Work
References
A Management Framework for Service Personalization
1
Introduction
2
Related Work
2.1
Content Service Network
2.2
Open Pluggable Edge Services Network
3
Service Personalization
4
Management Framework for Service Personalization
4.1
Policy Rule Modules
4.2
Service Manager Interaction
4.3
Content Access and Interaction with AAA Services
4.4
Subscriber Management
4.5
Fault Management
4.6
Service Accounting and Content Provider Interaction
5
Service Management Scalability and Performance
6
Conclusion
References
Efficient Access Using Hierarchical WML Decks for Multimedia Services under Wireless and Mobile Networks
1 Introduction
2 HTML Filter with Hierarchical Decks
2.1 Creation of Hierarchical Decks
2.2 Transmission of Hierarchical Decks in a Low Bandwidth Wireless Environment
2.2.1 Demand Driven Transmission (DDT)
2.2.2 WAP Protocol to Support the Demand Driven Transmission (DDT)
3 The Implementation of HTML Filter with Hierarchical Decks
3.1 RuleSet Database and Scripts
3.2 Parsing Engine (PE)
3.3 Markup Language Translator (MLT)
4 Experimental Results
4.1 Functional Evaluation
4.2 Performance Evaluation
5 Conclusions
Acknowledgment
References
Roaming Scenarios Based on SIP
1 Introduction
2 The Session Initiation Protocol
2.1 SIP Overview
2.2 SIP Registration Mechanism
2.3 Mobility Support with SIP
3 SIP Roaming Model
3.1 Reference Architecture and Logical Entities
4 Registration Scenarios for Roaming Users
4.1 Basic Home Registration
4.2 Third-Party Local Registration
a) Proxy Initiated
b) User Initiated
c) Home Initiated
4.3 Dual Registration
5 Conclusions
References
Low-Weight Congestion Control for Multi-sender Applications
1 Introduction
2 Requirements and General Assumptions
2.1 Sender-Based or Receiver-Based
2.2 A Potential Conflict
3 Related Work
4 Description of the Prototype
4.1 Quick NACK Feedback from the Worst Receiver
4.2 Selection of the Worst-Receiver
An Alternative to Receiver Reports
4.3 Configurablity
4.4 Initial Testing
4.5 Quality of Feedback
5 Summary and Future Work
References
Routing-Based Video Multicast Congestion Control
1
Introduction and Previous Work
2
The RVMCC Protocol
2.1
Layer Definition and Layer Priority in RVMCC
2.2
Solving Congestion and Claiming Bandwidth
2.3
Achieving Fairness and Stability
2.4 Further Design for Stability
2.5
Network Cost
3
Simulation Results
3.1
Network Topology and Parameters in Our Simulations
3.2
Simulation Scenarios
4
Summary
References
Random Early Detection Assisted Layered Multicast
1
Introduction
2
Loss Differentiation and Layered Multicast
3
RED Assisted Layered Mutlicast
3.1
Two-Level Drop Priority Support in IP Networks
3.2
Receiver-Driven Congestion Control Algorithm
4
Evaluations
4.1
Accuracy and Stability
4.2
Intra-session Interactions
4.3
Inter-session Interactions
5
Conclusion
References
Erratum to: Management of Multimediaon the Internet
Author Index
Recommend Papers

Management of Multimedia on the Internet
 3540442715, 9783540442714

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

Lecture Notes in Computer Science Edited by G. Goos, J. Hartmanis, and J. van Leeuwen

2496

3

Berlin Heidelberg New York Barcelona Hong Kong London Milan Paris Tokyo

Kevin C. Almeroth Masum Hasan (Eds.)

Management of Multimedia on the Internet 5th IFIP/IEEE International Conference on Management of Multimedia Networks and Services, MMNS 2002 Santa Barbara, CA, USA, October 6-9, 2002 Proceedings

13

Series Editors Gerhard Goos, Karlsruhe University, Germany Juris Hartmanis, Cornell University, NY, USA Jan van Leeuwen, Utrecht University, The Netherlands Volume Editors Kevin C. Almeroth University of California at Santa Barbara Department of Computer Science, Santa Barbara, CA 93106, USA E-mail: [email protected] Masum Hasan Cisco Systems Inc. 170 West Tasman Drive, San Jose, CA 95134, USA E-mail: [email protected]

Cataloging-in-Publication Data applied for Die Deutsche Bibliothek - CIP-Einheitsaufnahme Management of multimedia on the Internet : proceedings / 5th IFIP/IEEE International Conference on Management of Multimedia Networks and Services, MMNS 2002, Santa Barbara, CA USA, October 6 - 9, 2002 / Kevin C. Almeroth ; Masum Hasan (ed.). - Berlin ; Heidelberg ; New York ; Hong Kong ; London ; Milan ; Paris ; Tokyo : Springer, 2002 (Lecture notes in computer science ; Vol. 2496) ISBN 3-540-44271-5

CR Subject Classi cation (1998): C.2, H.5.1, H.3, H.5, K.3 ISSN 0302-9743 ISBN 3-540-44271-5 Springer-Verlag Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, speci cally the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on micro lms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are liable for prosecution under the German Copyright Law. Springer-Verlag Berlin Heidelberg New York a member of BertelsmannSpringer Science+Business Media GmbH http://www.springer.de © IFIP International Federation for Information Processing 2002 Printed in Germany Typesetting: Camera-ready by author, data conversion by Markus Richter, Heidelberg Printed on acid-free paper SPIN: 10870716 06/3142 543210

The original version of the book was revised: The copyright line was incorrect. The Erratum to the book is available at DOI: 10.1007/978-3-540-45812-8_28

Preface

The 5th IFIP/IEEE International Conference on Management of Multimedia Networks and Services (MMNS) was held in Santa Barbara, California. This was the fifth year that MMNS brought together researchers and developers to discuss the latest advances and explore future directions for managing multimedia in the Internet. As in past years, MMNS continues to be a competitive conference, attracting excellent papers with some of the top new ideas. This year we received 76 papers, of which 27 were accepted for inclusion in the program. What was particularly impressive this year was the large percentage of very high quality papers. The submissions made the job of the program committee extremely difficult. The span of topics this year ranged from network-layer traffic differentiation to application-layer consideration for multimedia traffic. At the network layer, a number of papers attempt to develop better solutions for differentiated services in the Internet. The issue of bandwidth sharing, particularly in wireless networks, is the focus of another set of papers. Next, by increasing the level of abstraction slightly, researchers are focusing on managing “services”. In particular, these services include one-to-many communication and video distribution architectures. Finally, from the user perspective there are two problems. What can applications do to help achieve better quality from the Internet? And, how can technology be applied to enterprise management systems? Taking the papers in the MMNS program together, they offer a range of solutions to key problems in managing multimedia traffic in the Internet. The success of MMNS can largely be attributed to high caliber committee members who worked hard to make the conference the best it could be. Each paper submitted was typically sent to four reviewers. After all the reviews were collected, papers with conflicting reviews were extensively discussed by committee members in an attempt to reach a consensus and fully examine the contributions of the work. The result is a program consisting of numerous outstanding papers. As conference chairs, it was a delight to work with such a dedicated and conscientious program committee. Finally, this year sees the continued support of IFIP and IEEE. Once again they handled many of the administrative tasks to make the conference run as smoothly as possible. In addition, we would like to thank Cisco Systems, Inc. for their financial support of MMNS. For those of you who attended the conference we hope you found this year’s MMNS to be a truly valuable experience, making new friends, visiting with old colleagues, and developing new ideas for next year’s MMNS. October 2002

Kevin Almeroth, Masum Hasan Conference Chairs

MMNS 2002 Organizing Committee

Steering Committee Ehab Al-Shaer, DePaul University Raouf Boutaba, University of Waterloo Giovanni Pacifici, IBM Research Guy Pujolle, University of Pierre & Marie Curie

Conference Co-chairs Kevin Almeroth, University of California, Santa Barbara Masum Hasan, Cisco Systems

Proceedings Chair Srinivasan Jagannathan, University of California, Santa Barbara

Web Chairs Robert Chalmers, University of California, Santa Barbara Sami Rollins, University of California, Santa Barbara

Organization

Program Committee

Kevin Almeroth, University of California, Santa Barbara Masum Hasan, Cisco Systems Nazim Agoulmine, University of Evry Ehab Al-Shaer, DePaul University Nikos Anerousis, VoiceMate.com Mohammed Atiquzzaman, University of Oklahoma Supratik Bhattacharryya, Sprint ATL Raouf Boutaba, University of Waterloo Greg Brewster, DePaul University Andrew Campbell, Columbia University Russ Clark, Georgia Tech Metin Feridun, IBM Research Dominique Gaiti, University of Troyes Mohsen Guizani, University of West Florida Abdelhakim Hafid, Telcordia Go Hasegawa, Osaka University Ahmed Helmy, University of Southern California David Hutchison, Lancaster University Muhammad Jaseemuddin, Ryerson Polytechnic Gautam Kar, IBM Research Ahmed Karmouch, University of Ottawa Lundy Lewis, APRISMA Derong Liu, University of Illinois, Chicago Songwu Lu, University of California, Los Angeles Hanan Lutfiyya, University of Western Ontario Allen Marshall, Queen’s University, Belfast Ahmed Mehaoua, University of Versailles Jose Neuman de Souza, University Fed. do Ceara Jose M. Nogueira, University Minas Gerais Giovanni Pacifici, IBM Research Guy Pujolle, University Pierre & Marie Curie Ed Perry, HP Labs Puneet Sharma, HP Labs Chien-Chung Shen, University of Delaware Rolf Stadler, Columbia University Ralf Steinmetz, University of Darmstadt Burkhard Stiller, UniBw Munich and ETH Zurich John Vicente, Intel Alaa Youssef, IBM Research

IX

X

Organization

Additional Reviewers

Nadjib Achir, LIP6 Toufik Ahmed, University of Versailles Chedley Aouriri, Intel Corporation Hakima Chaouchi, LIP6 Louise Crawford, Queen’s University, Belfast John Cushnie, Lancaster University Paulo-Andre Da-Silva-Goncalves, LIP6 Marcelo Dias de Amorim, LIP6 Martin Dunmore, Lancaster University Christopher Edwards, Lancaster University Idir Fodil, 6WIND Anelise Munaretto Fonseca, LIP6 Dan Gavenda, DePaul University Alberto Gonzales, CTR/Columbia University Jan Gerke, ETH Zurich Gerard Gross, Intel Corporation Hazem Hamed, DePaul University Hasan, ETH Zurich Bassam Hashem, Nortel Networks David Hausheer, ETH Zurich Pascal Kurtansky, ETH Zurich Rui Lopes, Lancaster University Dimitris Pezaros, Lancaster University Nicholas Race, Lancaster University Priya Rajagopal, Intel Corporation Govindan Ravindran, Soma Networks Abdallah Rayhan, Ryerson University Lopa Roychoudhuri, DePaul University Stefan Schmid, Lancaster University Nabil Seddigh, Tropic Networks Sakir Sezer, Queen’s University, Belfast Paul Smith, Lancaster University Steven Simpson, Lancaster University Perry Tang, Intel Corporation Yongning Tang, DePaul University Nguyen Thi Mai Trang, LIP6 Jason Yoa, DePaul University Bin Zhang, DePaul University

Table of Contents

Service Management A QoS Network Management System for Robust and Reliable Multimedia Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shirshanka Das, Kenshin Yamada, Heeyeol Yu, Scott S. Lee, Mario Gerla (University of California, Los Angeles, USA)

1

Federated Accounting Management System Architecture for Multimedia Service Usage Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Bharat Bhushan, Thomas Gringel (Fraunhofer-FOKUS, Germany), Conor Ryan, Eric Leray, Eamonn de Leastar, James Cloney (TSSG, Ireland) Policy-Based Quality of Service and Security Management for Multimedia Services on IP Networks in the RTIPA Project . . . . . . . . . . . . . . . . . . . . . . . . . 25 Val´erie Gay, Sandrine Duflos, Brigitte Kervella, Eric Horlait (Laboratoire d’Informatique de Paris 6, France), Gladys Diaz (L2TI - Institut Galil´ee - Universit´e Paris 13, France) Formal Modeling of Service Session Management . . . . . . . . . . . . . . . . . . . . . . . 36 Minh van Le, B.J.F. van Beijnum, B.L. de Goede (University of Twente, The Netherlands)

Management of Wireless Multimedia Network Requirement for Management of Multimedia over Wireless Channel 49 Bing Zheng (New Focus, USA), Mohammed Atiquzzaman (University of Oklahoma, USA) Agile Systems Manager for Enterprise Wireless Networks . . . . . . . . . . . . . . . 62 Sandeep Adwankar, Venu Vasudevan (Motorola Labs, USA) System Performance of HiperLAN/2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 Kamiran Haider, Hamed S. Al-Raweshidy (University of Kent at Canterbury, England)

Bandwidth Sharing Protocols Streaming Media Congestion Control Using Bandwidth Estimation . . . . . . . 89 Nadeem Aboobaker, David Chanady, Mario Gerla, M.Y. Sanadidi (University of California, Los Angeles, USA) Signaling Protocol for Session-Aware Popularity-Based Resource Allocation . 101 Paulo Mendes, Henning Schulzrinne (Columbia University, USA), Edmundo Monteiro (University of Coimbra, Portugal)

XII

Table of Contents

Distributed Video Architectures A Scalable Video Server Using Intelligent Network Attached Storage . . . . . 114 Guang Tan, Hai Jin, Liping Pang (Huazhong University of Science and Technology, Wuhan, China) On Proxy-Caching Mechanisms for Cooperative Video Streaming in Heterogeneous Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 Naoki Wakamiya, Masayuki Murata, Hideo Miyahara (Osaka University, Japan)

Management Systems Using CORBA’s Advanced Services to Enhance the Integrity of QoS Management Programmable Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 Qiang Gu, Alan Marshall (The Queen’s University of Belfast, Ireland) Remote Multicast Monitoring Using the RTP MIB . . . . . . . . . . . . . . . . . . . . . 154 Julian Chesterfield, Bill Fenner, Lee Breslau (AT&T Labs—Research, USA) Active Technology as an Efficient Approach to Control DiffServ Networks: The DACA Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 Nadjib Achir, Nazim Agoulmine, Mauro Fonseca, Yacine Ghamri-Doudane, Guy Pujolle (Pierre et Marie Curie University, Paris, France) Efficient Implementation of Packet Scheduling Algorithm on High-Speed Programmable Network Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 Weidong Shi, Xiaotong Zhuang, Indrani Paul, Karsten Schwan (Georgia Institute of Technology, USA)

Differentiated Network Services Distributed Dynamic Capacity Contracting: A Congestion Pricing Framework for Diff-Serv . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 Murat Yuksel, Shivkumar Kalyanaraman (Rensselaer Polytechnic Institute, Troy, USA) Fair Stateless Aggregate Traffic Marking Using Active Queue Management Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 Abhimanyu Das, Debjyoti Dutta, Ahmed Helmy (University of Southern California, USA) A Dynamic Marking Scheme of Assured Service for Alleviating Unfairness among Service Level Agreements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 Seung-Joon Seok, Seok-Min Hong, Chul-Hee Kang (Korea University, Korea), Sung-Hyuck Lee (Samsung Advanced Institute of Technology, Korea)

Table of Contents

XIII

Minimizing Transmission Costs through Adaptive Marking in Differentiated Services Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 Chen-Khong Tham, Yong Liu (National University of Singapore, Singapore) Dynamic QoS Adaptation Using COPS and Network Monitoring Feedback 250 Toufik Ahmed, Ahmed Mehaoua (University of Versailles, France), Raouf Boutaba (University of Waterloo, Canada)

User Level Traffic Adaptation Design and Implementation of an Application Layer Protocol for Reducing UDP Traffic Based on User Hints and Policies . . . . . . . . . . . . . . . . . . . . . . . . . 263 William Kulju (IBM Toronto Laboratory, Canada), Hanan Lutfiyya (University of Western Ontario, Canada) A Management Framework for Service Personalization . . . . . . . . . . . . . . . . . . 276 Govindan Ravindran, Muhammad Jaseemudin, Abdallah Rayhan (Ryerson University, Toronto, Canada) Efficient Access Using Hierarchical WML Decks for Multimedia Services under Wireless and Mobile Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 Dae-gun Kim (Korea Telecom, Seoul), Seung-Jin Lee (LG Electronic Inc, Anyang, Korea), Lynn Choi, Chul-Hee Kang (Korea University, Seoul, Korea) Roaming Scenarios Based on SIP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 Andrea Floris, Luca Veltri (CoRiTeL, Rome, Italy)

Multicast Congestion Control Low-Weight Congestion Control for Multi-sender Applications . . . . . . . . . . . 315 Jeremiah Scholl, Peter Parnes (Lule˚ a University of Technology, Sweden) Routing-Based Video Multicast Congestion Control . . . . . . . . . . . . . . . . . . . . 328 Jun Peng, Biplab Sikdar (Rensselaer Polytechnic Institute, Troy, USA) Random Early Detection Assisted Layered Multicast . . . . . . . . . . . . . . . . . . . 341 Yung-Sze Gan, Chen-Khong Tham (National University of Singapore, Singapore) Erratum to: Management of Multimedia on the Internet . . . . . . . . . . . . . . . . E1 Kevin C. Almeroth and Masum Hasan

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355

A QoS Network Management System for Robust and Reliable Multimedia Services S. Das, K. Yamada, H. Yu, S. S. Lee, and M. Gerla Computer Science Department University of California, Los Angeles Los Angeles, CA 90095-1596 {shanky, kenshin, heeyeoly, sslee, gerla}@cs.ucla.edu Abstract. In this paper, we introduce a practical network management system with QoS for multimedia traffic. Today, multimedia applications have evolved significantly and have become an essential part of the Internet. In order to effectively support the newly emerging network traffic, the underlying network protocols need to be aware of the characteristics and demands of the traffic. The proposed network system is ready to address the varied characteristics of multimedia traffic and can assure a high degree of adherence to the quality of service demanded from it. Considering the need for reliable QoS services, the system is also equipped with fault tolerance capability by provisioning multiple QoS paths. Moreover, the system has been practically designed and implemented to provide “cost-effective” QoS support with respect to control overhead. It deploys measurement-based QoS path computation and call admission scheme which may be deemed ineffective for bursty multimedia traffic. However, results conclusively prove that provisioning multiple paths and utilizing them in parallel in our system not only provides high fault tolerance capability but also effectively accommodates multimedia traffic by relieving its burstiness with multiple paths. We present the architecture of the system and discuss the benefits gained for multimedia traffic.

1

Introduction

Recently, the Internet has become a significant medium for real-time data as demands for multimedia services become more and more pervasive. In order to accommodate such new types of traffic in the Internet, researchers have addressed various issues with respect to provisioning QoS. QoS provisioning issues consist of several building blocks such as resource assurance, service differentiation, and routing strategies. Among these, QoS routing as a routing strategy is essential since it executes computations to find QoS-feasible paths. [1,2] address a possible use of the Bellman-Ford algorithm in the link state (OSPF) routing environment for QoS routing. In addition, [3,4] showed specifications and feasibility of such QoS routing approaches. The development of the Multiprotocol Label Switching Protocol (MPLS) [5] and the increased acceptance of Traffic Engineering as an important component of routing protocols has accelerated the coming of age of QoS routing. The original version of this chapter was revised: The copyright line was incorrect.This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 1–11, 2002. c IFIP International Federation for Information Processing 2002

2

S. Das et al.

In this paper, we present a practical QoS network management system based on the proposed OSPF-based QoS architecture. This consists of OSPF with traffic engineering extensions running as the routing protocol, MPLS for explicitly routing packets, and a QoS path provisioning algorithm that also serves as a fundamental call admission control (CAC). The proposed network system is equipped with fault tolerance capabilities for unreliable network environment. This fault tolerance issue in QoS provisioning is considered as a very important aspect especially when mission-critical applications are running. This issue has been highlighted in [6,7]. Accordingly, a novel approach was recently introduced in [8] which deploys an effective algorithm to produce multiple QoS paths. Multiple QoS path provisioning also showed various benefits such as even network resource utilization and cost-effective link state information acquisition [9]. In addition, our proposed system in this paper provides statistical QoS guarantees. That is, it deploys measurement-based path computations and call admissions without any resource reservations. This helps in reducing the system complexity and achieving relatively fast QoS provisioning. It is well known that multimedia applications are likely to produce highly bursty traffic with hardly predictable characteristics [10,11]. This high burstiness keeps measurement-based approaches from properly performing. This is because the severe traffic fluctuation misleads the approaches into producing non-feasible QoS paths and making incorrect call admission decisions. The system that we propose carries out rather reliable call admissions and performs well since it provisions multiple QoS paths and spreads network traffic over them. Thus, it diffuses traffic burstiness and attenuates its negative impact. Targeting such a fault-tolerant and reliable QoS-service architecture, we have implemented a QoS testbed. In this paper, we present the system architecture with the QoS routing approach for multiple path and show the fault tolerance capability with reliable multimedia services. In Section 2, the routing algorithms are briefly reviewed. Section 3 depicts the entire network system architecture, and Section 4 presents experiment results obtained with the implemented network system and demonstrates its effectiveness in representative traffic scenarios.

2

QoS Algorithms

The network system presented in this paper has QoS routing mechanisms which are ready to serve QoS applications in both conventional and fault-tolerant ways. The mechanisms serve as a simple call admission control (CAC). When a QoS application comes in and looks for its corresponding QoS services, it consults the underlying QoS routing algorithm (i.e., Q-OSPF with the enhanced routing algorithms in our case) for feasible paths. If no feasible path is found, the connection is rejected and the application exits. Thus, the QoS routing path computation algorithm not only provides the capability of finding QoS paths but also plays an important role in CAC. As previously mentioned and discussed in [8], we adopted the two different QoS routing algorithms in our system; the conventional single path algorithm and the newly introduced multiple path algorithm.

A QoS Network Management System

3

The single QoS path computation algorithm with multiple QoS constraints derives from the conventional Bellman-Ford algorithm as a breadth-first search algorithm minimizing the hop count and yet satisfying multiple QoS constraints. Each node in the network builds the link state database which contains all the recent link state advertisements from other nodes. With Q-OSPF, the topological database captures dynamically changing QoS information. The link state database accommodates all the QoS conditions, and we define each condition as a QoS metric and each link in the network is assumed to be associated with multiple QoS metrics which are properly measured and flooded by each node. Each of these QoS metrics has its own properties when operated upon in the path computation. The principal purpose of the path computation algorithm is to find the shortest (i.e., min-hop) path among those which have enough resources to satisfy given multiple QoS constraints, rather than the shortest path with respect to another cost metric (e.g., maximizing available bandwidth or minimizing end-to-end delay). The proposed multiple QoS path algorithm is a heuristic solution. We do not limit ourselves to strictly “path disjoint” solutions. Rather, the algorithm searches for multiple, maximally disjoint paths (i.e., with the least overlap among each other) such that the failure of a link in any of the paths will still leave (with high probability) one or more of the other paths operational. The multiple path computation algorithm can then be derived from the single path computation algorithm with simple modifications. This multiple path computation algorithm produces incrementally a single path at each iteration rather than multiple paths at once. All the previously generated paths are kept into account in the next path computation. The detailed descriptions of the single and the multiple path algorithms are in [8].

3

System Architecture

The testbed consists of PCs running Linux, and all the QoS-capable features are embedded in the Linux kernel. Each of the machines has several modules running on it, namely the link emulator, metric collector, OSPF daemon, MPLS forwarding and the applications, and the entire system architecture is depicted in Fig. 1. The following sections describe in more detail each of these modules individually, explaining their functions and their implementation issues. 3.1

Link Emulator

The testbed that we implement uses the wired Ethernet LAN in the lab, as the physical communication medium. Since we want to emulate several scenarios with widely varying link capacities and propagation delays, we require a network emulator to do the job for us. We emulate the characteristics of bandwidth and delay over a link using tc, the Linux kernel traffic shaper and controller. tc can emulate a wide variety of policies with hierarchies of filters and class based queues on the outgoing interfaces. We require simple bandwidth and delay emulation

4

S. Das et al. CBR Generator

MPEG4 Player (Client)

MPEG4 Streaming Server

: LSA Flooding : QoS Path Computation

: Measuring QoS Metrics

Q−OSPF Daemon

Metric Collector

Link Emulator

Linux Kernel with MPLS

Fig. 1. The entire QoS system architecture for the experiment testbed. which can be done quite easily. A simple “tc qdisc add dev eth0 root tbf rate 10kbps latency 10ms” emulates a 10 Kbps link with a latency of 10 ms on the eth0 interface. Since we have bursty multimedia traffic, the buffer size is also an important factor, and after a few experiments, we settled on a value of 25 Kbits as a reasonable buffer to accommodate burstiness. 3.2

Metric Collector

To provide QoS, reserving bandwidth and bounding on delay for a connection, we require the knowledge of link characteristics at all times. This information is needed by the QoS allocation module which looks at the current usage of the network and figures out if the new request can be satisfied or not. The link metric collection module, therefore, is an integral part of any routing scheme implementation that provides QoS routing. The OSPFD (OSPF daemon) implementation does not have any link characteristics measurement module. The metrics we are interested in are the available bandwidth in the link and the delay. One could think of other metrics also, such as queue length, loss probability, etc., but bandwidth and delay are the two most important metrics used in QoS routing. Thus, our design goal was to write a module to measure these two metrics and integrate this code with the existing OSPFD implementation. For bandwidth metric collection, we opted to simply use the log file maintained in /proc/net/dev/ which contains information about the number of packets and bytes received and transmitted on each interface. By examining this file at regular intervals, we calculate the bandwidth used on the each outgoing interface. Delay metric collection is done using the ‘ping’ utility to send ping probes to the other side of the link and collect the delay value. The metric collection code sends ping message and collects bandwidth values from the /proc/net/dev log file, every time interval (typically 10 ms). The values collected are exponentially averaged to smooth out the fluctuations. 3.3

Q-OSPF Daemon

To propagate QoS metrics among all routers in the domain, we need to use an Interior Gateway Protocol (IGP). OSPF is one of the major IGPs and significant

A QoS Network Management System

5

researches have been recently made on OSPF with traffic engineering extensions. We selected the open source OSPF daemon (OSPFD) [12,13] to implement our QoS routing scheme. [14] defines Opaque LSA for OSPF nodes to distribute user-specific information. Likewise, we define our specific Opaque LSA entries by assigning new type values in the Opaque LSA format shown in Fig. 2. When OSPFD runs at routers, it tries to find its neighbor nodes by sending HELLO messages. After establishing neighbor relationship, OSPFD asks the metric measurement module to calculate the QoS metrics of the established link. OSPFD gives as input the local interface address and the neighbor router interface address to the metric measurement module and generates the opaque LSA for each interface. The LSA contains the available bandwidth and queuing delay metric obtained from the metric measurement module. LSA update frequency is currently restricted to MinLSInterval (5 seconds). In our implementation, we followed this restriction, but we feel that the LSA update interval is an important parameter for the correctness of our algorithms because if there are too frequent changes to the generated LSA, there will be miscalculations during the QoS path computation due to the flooding delay. In addition to LSA flooding, OSPFD exchanges router LSAs to build a full network topology. Router LSAs originate at each router and contain information about all router links such as interface addresses and neighbor addresses. We bind the link metrics that we are interested in, viz. bandwidth and delay to the opaque LSA specified by the link interface address. Thus, we introduce the pointer from each link of a router LSA to the corresponding opaque LSA as shown in Fig. 3. Whenever OSPFD receives router LSAs or opaque LSAs, it just updates the link pointer to point to the new traffic metrics reported by that link. LS age Opaque type

Options Opaque ID Advertising router

LS type 10

LS sequence number LS checksum Length Type = 2 Length = 32 Type = 1 Length = 1 All 0 P2P = 1 Type = 4 Length = 4 Local interface address Type = 32768 Length = 4 Available bandwidth (bps) IEEE floating point Type = 32769 Length = 4 Delay (sec) IEEE floating point

Link state header

D Link3 Link TLV Link type sub TLV Local interface IP address sub TLV Available bandwidth sub TLV Link delay sub TLV

Fig. 2. Opaque LSA format.

3.4

Router LSA B

A Link1 B

Link2 C

Host A Link1 info Link2 info Link3 info

Router LSA C Router LSA D

Link1 Opaque LSA AV BW

Link2 Opaque LSA AV BW

Link3 Opaque LSA AV BW

Delay

Delay

Delay

Opaque LSA Opaque LSA Opaque LSA Opaque LSA Opaque LSA Opaque LSA

Fig. 3. Router LSAs.

MPLS

One of the key assumptions for Q-OSPF to be effective is the capability of setting up explicit paths for all packets of a stream to use. Thus, we need to pin down the

6

S. Das et al.

path that a connection uses. One of the main advantages of MPLS is its efficient support of explicit routing through the use of Label Switched Paths (LSPs). With destination-based forwarding as in the conventional datagram networks, explicit routing is usually provided by attaching to each packet the network-layer address of each node along the explicit path. This approach makes the overhead in the packet prohibitively expensive. This was the primary motivation behind deploying MPLS at the lower layer instead of using something like IP sourcerouting or application level forwarding. 3.5

Applications

An integral part to the whole process of making a testbed was the development of applications running on top of this architecture. These applications request QoS-constrained routes, generate traffic, and in turn change the QoS metrics of the network. We used the open source Darwin Streaming Server distributed by Apple [15] to stream MPEG-4 files using RTSP over RTP. The open source mp4player from the MPEG4IP project [16] was used to play these files over the network. Modifications to the source code involved extending the capabilities of the server and the player for streaming and playing over multiple paths, and designing an interface for the server to interact with Q-OSPF daemon and request paths to the client. For the experiments involving CBR traffic, we used a home-brewed traffic generator which generates UDP traffic at a given rate in either single path or multiple path mode.

4

Experiments

We performed two kinds of experiments for this paper. The first set of experiments involves comparing the performance of Q-OSPF with respect to the throughput of multimedia traffic using single paths and using multiple paths. The second set involves comparing the performance of Q-OSPF with respect to fault tolerance and robustness of multimedia connections using single path and using multiple path. Fig. 4 shows the network topology for the experiments. 9 nodes (i.e., routers) are connected directly to each other through 1.5 Mbps links. We performed experiments in certain representative scenarios where multiple path is expected to give better results as well as in totally general scenarios. 1. The first experiment involved sending six streams of multimedia traffic from QOS6 to QOS5. There are three disjoint paths available that satisfy the bandwidth constraints, QOS6 - QOS2 - QOS1 - QOS5, QOS6 - QOS7 QOS4 - QOS5, QOS6 - QOS9 - QOS8 - QOS5. Let us name them QPath1, QPath2 and QPath3 respectively. We run the experiment in both Single Path and Multi Path mode. In single path mode, each stream uses one of the three QPaths, while in multi path mode, each stream uses all the three QPaths. Results are in Fig. 5 and Fig. 6.

A QoS Network Management System QoS1

7

QoS2

00 11 11 00

00 11 11 00 QoS3

00 11 11 00 QoS5

QoS4

00 11 11 00

QoS7

00 11 11 00

QoS6

00 11 11 00

QoS8

00 11 11 00 QoS9

00 11 11 00

00 11 11 00

Fig. 4. The QoS testbed topology 1400

Link 1 Link 2 Link 3

1200

1200

1000

1000 Traffic Bandwidth (Kbps)

Traffic Bandwidth (Kbps)

1400

800

600

800

600

400

400

200

200

0

0

20

40

60 Time (sec)

80

100

120

Link 1 Link 2 Link 3

0

0

20

40

60 Time (sec)

80

100

120

Fig. 5. Throughput profile on the three Fig. 6. Throughput profile on the three QPaths: single path

QPaths: multi path

2. The second experiment involved sending three CBR streams of 800 Kbps first on the three QPaths, and then introducing the six multimedia streams later. In the single path case, the calls get accepted and follow the same paths as before, two streams each sharing one QPath, but in this case due to burstiness, the total throughput on each path frequently goes above the 1.5 Mbps limit, and thus there are losses. One the other hand, with multi path mode, the traffic on all the paths is much smoother, and thus the total traffic never goes above 1.5 Mbps on any link and thus there are no losses. This scenario demonstrates the benefits on multi path with respect to smoothening the bursty traffic. Results are in Fig. 7 and Fig. 8. 3. The last experiment in this set was a totally general scenario, in which random connections were set up, and the traffic characteristics were measured. We first perform the experiment for well behaved CBR traffic in the single path mode and then perform it with bursty multimedia traffic first using single path mode and then multiple path.

8

S. Das et al. 1800

1400

1400

1200

1200

1000 800 600 400

1000 800 600 400

200 0

Link 1 Link 2 Link 3

1600

Traffic Bandwidth (Kbps)

Traffic Bandwidth (Kbps)

1800

Link 1 Link 2 Link 3

1600

200

0

20

40

60

80 Time (sec)

100

120

140

160

0

0

20

40

60

80

100

120

140

Time (sec)

Fig. 7. Throughput profile on the three Fig. 8. Throughput profile on the three QPaths: single path

QPaths: multi path

Owing to performance bounds of the multimedia server and the CPUs of the nodes, we reduced the bandwidth emulated on each link to 1 Mbps. The total number of connections generated in both CBR and multimedia case was 28. These connections were generated using randomly chosen source and destination from the 9 QoS routers. The measured traffic characteristics are the usual throughput profile on the links. Fig. 9 shows the throughput characteristics on each link when CBR traffic is in the network and Q-OSPF is in single path mode. Fig. 10 shows the throughput characteristics when multimedia traffic (of the same average bandwidth as the CBR in the previous experiment) is in the network and Q-OSPF is in single path mode. We can see that with multimedia the performance of Q-OSPF is visibly worse and the traffic frequently goes above the threshold of 1 Mbps. Fig. 11 shows the throughput characteristics when multiple paths are provisioned. Here we see that on each link the burstiness is much less and the overall QoS constraints are being satisfied. The throughput is below the 1 Mbps limit. Thus, we clearly see the benefits of using multiple paths even in a completely random general scenario. The second set of experiments examines the fault tolerance capability by provisioning multiple QoS paths between source and destination and spreading packets over the multiple paths. The network topology is the same as in Fig. 4. Here again we experiment with two extreme traffic scenarios, well behaved CBR traffic and bursty multimedia traffic. The pairs of source and destination nodes for the connections are randomly selected, and the link failures occur also randomly during the execution of the experiments. For the link failure model, we used a exponential distribution such that the average link down time is equal to 5 seconds. The number of multiple paths requested is kept at 3. The statistics which we collect during this experiment is the amount of time for which each connection is down. A connection is said to be down when none of its packets are reaching the destination for 2 seconds.

A QoS Network Management System 1400

QoS2 - QoS1 QoS3 - QoS1 QoS5 - QoS1 QoS1 - QoS2 QoS3 - QoS2 QoS1 - QoS3 QoS2 - QoS3 QoS4 - QoS3 QoS7 - QoS3 QoS3 - QoS4 QoS5 - QoS4 QoS7 - QoS4

1200

Throughput (Kbps)

1000 800

QoS1 - QoS5 QoS4 - QoS5 QoS5 - QoS5 QoS2 - QoS6 QoS7 - QoS6 QoS3 - QoS7 QoS4 - QoS7 QoS6 - QoS7 QoS4 - QoS8 QoS5 - QoS8 QoS6 - QoS9 QoS7 - QoS9

1000

600 400

800

600

400

200 0

QoS2 - QoS1 QoS3 - QoS1 QoS5 - QoS1 QoS1 - QoS2 QoS3 - QoS2 QoS1 - QoS3 QoS2 - QoS3 QoS4 - QoS3 QoS7 - QoS3 QoS3 - QoS4 QoS5 - QoS4 QoS7 - QoS4 QoS1 - QoS5 QoS4 - QoS5 QoS5 - QoS5 QoS2 - QoS6 QoS7 - QoS6 QoS3 - QoS7 QoS4 - QoS7 QoS6 - QoS7 QoS4 - QoS8 QoS5 - QoS8 QoS6 - QoS9 QoS7 - QoS9

1200

Throughput (Kbps)

1400

9

200

0

20

40

60

80 100 Time (sec)

120

140

160

180

0

0

20

40

60

80 100 Time (sec)

120

140

160

180

Fig. 9. Throughput profile on every link: Fig. 10. Throughput profile on every CBR: Single Path

link: Multimedia: Single Path 1400

QoS2 - QoS1 QoS3 - QoS1 QoS5 - QoS1 QoS1 - QoS2 QoS3 - QoS2 QoS1 - QoS3 QoS2 - QoS3 QoS4 - QoS3 QoS7 - QoS3 QoS3 - QoS4 QoS5 - QoS4 QoS7 - QoS4 QoS1 - QoS5 QoS4 - QoS5 QoS5 - QoS5 QoS2 - QoS6 QoS7 - QoS6 QoS3 - QoS7 QoS4 - QoS7 QoS6 - QoS7 QoS4 - QoS8 QoS5 - QoS8 QoS6 - QoS9 QoS7 - QoS9

1200

Throughput (Kbps)

1000

800

600

400

200

0

0

20

40

60

80 100 Time (sec)

120

140

160

180

Fig. 11. Throughput profile on every link: Multimedia: Multi Path

The duration of the experiments was 10 minutes each. Fig. 12 and Fig. 13 show the connection down-time percentage when single path or multiple paths are provisioned for the connections for both the CBR traffic case and the multimedia case. The result shows that multiple path provisioning has lower connection down rate as expected. We can see that the benefits although almost equal are slightly lesser than in the CBR traffic case. The reason is that a lot of multimedia calls do not get 3 multiple paths due to the burstiness of existing connections which fill up the pipe. Thus, we show that we have greater fault tolerance for multimedia traffic when we provision multiple QoS paths compared to single QoS path. Note that the experiments are slightly different in spirit from the simulation experiments in [8]. In [8], the authors assume that a connection gets dropped when all its paths have one or more faults, and does not come up automatically when the fault gets repaired. However, the experiments have enough similarity to give similar results.

10

S. Das et al. 1

Single QoS Path provisioning Multiple QoS Path provisioning

0.1

Connection Down Rate

Connection Down Rate

1

0.01

0.001 0.1

0.2

0.3

0.4 0.5 Link Failure Rate

0.6

0.7

0.8

Single QoS Path provisioning Multiple QoS Path provisioning

0.1

0.01

0.001 0.1

0.2

0.3

0.4 0.5 Link Failure Rate

0.6

0.7

0.8

Fig. 12. Connection down rate as a func- Fig. 13. Connection down rate as a function of link failure rates: CBR tion of link failure rates: Multimedia

5

Conclusion

We proposed a QoS network management system in which QoS provisioning mechanisms are measurement-based and multiple QoS paths are computed. With the practical system implementation and experiments, we discussed the interaction of measurement-based call admission and bursty multimedia traffic. The experiment results verified that the bursty nature of multimedia traffic degrades the performance of measurement-based approaches. However, when multiple QoS paths are provisioned, the burstiness becomes diffused and its negative impact is lessened as the results showed. We also evaluated the fault tolerance capability of the network and found that the multiple paths provide a high degree of robustness in servicing QoS applications. In conclusion, provisioning multiple paths during call admission not only attenuates the impact of the burstiness but also provides highly robust QoS services.

References 1. R. Guerin, A. Orda, and D. Williams. QoS Routing Mechanisms and OSPF Extensions. In Proc. of Global Internet (Globecom), Phoenix, Arizona, November 1997. 1 2. Dirceu Cavendish and Mario Gerla. Internet QoS Routing using the Bellman-Ford Algorithm. In IFIP Conference on High Performance Networking, 1998. 1 3. G. Apostolopoulos, S. Kama, D. Williams, R. Guerin, A. Orda, and T. Przygienda. QoS Routing Mechanisms and OSPF Extensions. Request for Comments 2676, Internet Engineering Task Force, August 1999. 1 4. Alex Dubrovsky, Mario Gerla, Scott Seongwook Lee, and Dirceu Cavendish. Internet QoS Routing with IP Telephony and TCP Traffic. In Proc. of ICC, June 2000. 1 5. E. Rosen, A. Viswanathan, and R. Callon. Multiprotocol Label Switching Architecture. Request for Comments 3031, Internet Engineering Task Force, January 2001. 1

A QoS Network Management System

11

6. Shigang Chen and Klara Nahrstedt. An Overview of Quality of Service Routing for Next-Generation High-Speed Networks: Problems and Solutions. 12(6):64–79, November 1998. 2 7. Henning Schulzrinne. Keynote: Quality of Service - 20 Years Old and Ready to Get a Job? Lecture Notes in Computer Science, 2092:1, June 2001. International Workshop on Quality of Service (IWQoS). 2 8. Scott Seongwook Lee and Mario Gerla. Fault Tolerance and Load Balancing in QoS Provisioning with Multiple MPLS Paths. Lecture Notes in Computer Science, 2092:155–, 2001. International Workshop on Quality of Service (IWQoS). 2, 2, 3, 9, 9 9. Scott Seongwook Lee and Giovanni Pau. Hierarchical Approach for Low Cost and Fast QoS Provisioning. In Proc. of IEEE Global Communications Conference (GLOBECOM), November 2001. 2 10. S. Floyd. Comments on Measurement-based Admissions Control for Controlledload Services, 1996. 2 11. Lee Breslau, Sugih Jamin, and Scott Shenker. Comments on the Performance of Measurement-Based Admission Control Algorithms. In INFOCOM (3), pages 1233–1242, 2000. 2 12. J. Moy. OSPF Version 2. Request for Comments 2328, Internet Engineering Task Force, April 1998. 5 13. John T. Moy. OSPFD Routing Software Resources. http://www.ospf.org. 5 14. R. Coltun. The OSPF Opaque LSA Option. Technical report. 5 15. Apple Computer. The Darwin Streaming Server. http://www.opensource.apple.com/projects/streaming. 6 16. Bill May David Mackie and Alix M. Franquet. The MPEG4-IP Project. http://mpeg4ip.sourceforge.net. 6

Federated Accounting Management System Architecture for Multimedia Service Usage Management B. Bhushan1, T. Gringel1, C. Ryan2, E. Leray2, E. de Leastar2, and J. Cloney2 1

Fraunhofer-FOKUS, Kaiserin-Augusta-Allee 31, D-10589 Berlin Germany {bhushan, gringel}@fokus.fhg.de

2

TSSG, Ground Floor Unit A, Waterford Business Park, Cork Road Waterford Ireland {edeleastar,jcloney,cryan,eleray}@tssg.wit.ie

Abstract. Service consumers are increasingly becoming aware of QoS and service subscriptions. Service providers (SPs) are also recognizing the opportunity to generate revenue by bundling services in packages and providing them to customers. Standardization bodies have begun to address requirements of charging and usage accounting management as SPs are increasingly relying on them. Federated accounting management concept stems from these developments. It specifically supports the requirement of SPs to co-operate in a federated manner and share the revenue. It also supports the requirements of charging and billing being subcontracted to a third-party. This paper proposes that standard-based interfaces and shared information model, and a wellaccepted development methodology are keys to enabling co-operation between SPs. It presents a requirement analysis and a federated accounting management system architecture, which are based on the recommendations laid down by TMForum, IPDR, OMG and IETF organisations. The architecture supports mediation, rating, charges aggregation, settlement, and online billing.

1

Introduction

As broadband networks and UMTS infrastructure are more widely deployed, IP-based services are set to become more advanced and will be provided in customized manner. As a result, the demands of e-business are going to drive rapid innovation and SPs are going to provide new services in a B2B (business-to-business) environment. Recent surveys such as [1] and [2] chart these developments in a B2B context. The emergence of the B2B market is also influencing the way services are provided, charged and billed. An important feature in this development is the usage of composite services (several service bundled in one). Composite services consist of several types of services and can be provided by ISPs, virtual private network (VPN) and application service providers (ASP). In this environment, SPs face two important challenges: (1) operating services and maintaining QoS across multiple administrative domains; (2) integrating the charges for various services that make up a composite service. These challenges manifest themselves in a number of ways. For example, new types of B2B interactions have emerged mainly at the levels of service operations The original version of this chapter was revised: The copyright line was incorrect. This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 12-24, 2002. © IFIP International Federation for Information Processing 2002

Federated Accounting Management System Architecture

13

management. New requirements for customer support, Quality of Service (QoS) maintenance, charge settlement and billing have also emerged [3]. SPs have also acknowledged the emergence of a new stakeholder in the existing business model. The EU-sponsored IST project FORM in which the work presented in this paper was carried out termed this new stakeholder inter-enterprise service provider or IESP. The IESP can function as a portal to a wide array of services (simple or composite). It can also function as a service retailer or as a specialised service provider who composes new services on customersí demand and deal with the customer queries and bills on behalf of other SPs. This role of IESP adds a new dimension to Billing business model, which is, outsourcing of Billing operations management. Charging and billing is something without which SP businesses will not be commercially viable [4, [5]. Outsourcing Billing is proving to be effective because billing operations are increasingly becoming a specialised and automated task. The core management operations that SPs require involve customer support, QoS management, charging, billing, and SP settlement [13]. The IESP can play a crucial role in providing them where it deals with customer queries and SLAs (Service Level Agreements) on behalf of SP [6]. It can also do the settlement of the charges amongst SPs if they collectively provided the service and present the bill to the customer. Here one can see three functional areas in operation: (1) customer support and SLA management; (2) QoS management; (3) charging and billing. 1.1

Related Works

Organisations such as TMForum (Telemanagement Forum), IPDR (Internet Protocol Detail Record) and IETF, OMG (Object Management Group) and ETSI (European Telecommunication Standards Institute) have come up with guidelines to develop OSS in these three functional areas. The work presented in this paper is relevant to these organisations; hence a brief introduction to their works is given in this section. TMForum through its work on TOM (Telecom Operations Map) [7] and its successor eTOM [8] has produced a map of operations processes that can be used for the development of three functional areas mentioned above. Within TOM, these functional areas are called Fulfilment, Assurance and Billing (or FAB). The main goals of IPDR organisation are to define an open, flexible record format (the IPDR Schema) for exchanging usage information and essential parameters for any IP transaction [9]. For the specification of interfaces between OSSs (operation support system) and BSSs (business support system), the IPDR organisation has adopted the core functional roles and interfaces that TMFís TOM [7] has identified. The AAA (Authentication, Authorisation, Accounting) WG of IETF has also produced a set of RFC on accounting management protocol and services [10], [11]. The AAA group can be compared to TOM process flow-through where OSS is regarded as a complete process chain. In the chain Fulfilment feeds relevant information into Assurance, which in turn feeds relevant information into Billing (FAB). FAB themselves roughly correspond to Authentication, Authorisation and Accounting. Of these three broad areas, this paper focuses on the Billing area of TOM or Accounting sub-group of AAA WG. Having introduced the paper, an outline of the rest of the paper is given here. Section 2 presents the concept of the federated accounting management. Section 3

14

B. Bhushan et al.

describes the business context and presents requirement analysis model. Section 4 presents the federated accounting information model. Section 5 describes the federated accounting system architecture in details. Section 6 outlines the technology used to implement the system architecture. Section 7 concludes this paper and discusses the future work.

2

Federated Accounting Management Concept

This section introduces the concept of federated accounting management and briefly describes its main characteristics. Distribution of Operational Processes and Use of Interdomain Interfaces: Service mediation, rating, and billing processes are distributed over several administrative domains. These processes interoperate by the means of well-defined interfaces. Interfaces may also exist at the boundary between two domains, in which case they are also called Reference Point. The processes mentioned here are already widely known through the TOM map where they are parts of the Billing process [7]. Interfaces between processes have also become known through the work of IPDR Operational Model [9]. However interfaces between domains are not covered in sufficient details by TOM and IPDR and still require further investigation. This paper presents the design of interdomain interfaces (see Sect. 5). Inter-working of Fulfilment, Assurance and Billing Processes: These three processes take part in so-called end-to-end process flow. Although the Fulfilment and Assurance are separate functional areas in their own right, they co-operate with the Billing process. Federated accounting management incorporates this end-to-end process flow where Fulfilment and Assurance provide Billing with SLA and QoS details. This enables the service provider to apply different tariffs for different types of subscriptions or QoS provided. A More Inclusive View of Accounting Information: Federated accounting management holds a more inclusive view in which usage information and charges details together form a part of accounting information. IPDR organisation is working on a standardised information model. However, the work of IPDR is confined to producing specification of usage information exchanged. The concept of federated accounting management complements the work of IPDR organisation by providing rationale for and specification of charge details as a part of federated accounting information model (see Sect. 4). Use of Differential Tariffs: Charging regime uses tariffs that vary with customer types, QoS or any other factor that SPs choose to use. An earlier work by authors on this can be found in [12]. Charge Aggregation and Payment Settlement: In federated accounting management several third-party providers collectively provide their services. IESP acts on behalf of both customer and third-party providers. It aggregates charges and invoices the customer for all the services used; a single invoice is sent to the customer. IESP may also do the settlement for all third-party providers. Work on supplier-provider settlement is currently being addressed in eTOM [8].

Federated Accounting Management System Architecture

3

15

Business Context and Functional Requirements Analysis

This section presents the business context, actors involved in federated accounting management and the roles they play (see Fig. 1). The information exchanged between various roles is also shown. This is followed by a description of operations processes involved in federated accounting management. Thereafter key functional requirements that the architecture implements are presented as a use case diagram. IE S C u s to m e r

In t e r -E n t e r p r is e S e r v ic e (IE S ) T h ir d - P a r t y P r o v id e r P r o v id e r A g re e S L A IP D R s

M e d ia S h o p

Q u e r y /P a y B ill

A u th o r iz e

S e ttle m e n t B ro k e r

IP D R s

U s e rs U s e O n lin e C o lla b o r a t io n S e r v ic e ( O C S )

OCS USER

B 2C

S e ttle m e n t

V oIP G a te w a y

B 2B

Fig. 1. Federated Accounting Management Business Model

In the business context, the IESP ensures that the customer (or IES Customer in Fig. 1) has access to a variety of services that it needs. They are provided by many SPs (third-party provider in the Fig. 1), who are IESPís partners. OCS (online collaboration service) is one such service, which packages MediaShop and VoIP services in single application. MediaShop is multimedia contents distribution service. The IES Customer is the subscriber of IES and negotiates and signs a SLA with the IES Provider. It can be a business (organisation) or a private customer. It receives and pays bills for the services used. End-User is a person who uses Inter-Enterprise Service. End-user may be a private user or he/she may work for the IES Customer. The IESP, as the customer-facing SP, is responsible for SLA handling, tariff negotiation, customer care and accounting aspects of the services offered to the customer by the IESPís partners in a value chain. It provides the IES Customer with access to communication, application, and information services, which are in turn provided by third-party providers (i.e., IRS Provider). IES Provider performs the task of charging and billing of service usage and charge settlement among IES Customers and third-party SPs. IES Provider acts as a service retailer and maintains contracts with third-party providers and does the final payment settlement. The IRS (Information Retrieval Service) Provider is a third-party provider that provides its services to IES Customer. The third-party providers must be able to charge the IES Customer for service usage, details of which are sent to the IESP as IPDR documents (also shown Fig. 1). IESP incorporates the various IPDR documents by means of aggregation and rating processes and provides a consolidated bill to the

16

B. Bhushan et al.

IES Customer. In order for interaction to occur between the IESP and third-party providers, IPDR Master schema and specifications are used as a shared information schema and a set of interfaces that all parties support. 3.1

Operations Processes Involved in Federated Accounting Management

Fig. 2 illustrates the federated accounting management operations processes in the context of the FAB of TOM [7]. Fig. 2 is based on an original version of FAB figure [7]; various aspects of federated accounting management have been added to original FAB figure and illustrated in Fig. 2. An enhanced version of IPDR XML Schema [9] is used to model accounting information (or E-IPDRs shown in Fig. 2) that is exchanged between various business processes. Sect. 4 discusses the enhancement done to Master IPDR Schema. Subscription

Assurance Monitoring

Invoicing & Bill Presentment

Customer

4. Bill content

Bill Presentment and querying

Rating & Discounting 3. Aggregated E-IPDRs

Federated Mediation Adapter

Processes Fulfilment

Other SPs E-IPDRs

2. E- IPDRs

Assuranc N on - IESP

IPDR Mediator

1. Network, service, application (usage) data N etwork Element Management & Application Servers

Fig. 2. Federated Accounting Management Operations Process

3.2

Functional Requirements and Use Case Model

A comprehensive list of functional requirements is discussed in [12]. The authors have worked on those requirements and narrowed them down to a set of requirements that must be addressed in the design. The narrowed set includes the following requirements and the use case diagram (see Fig. 3) models them: 1. 2. 3. 4.

Usage-based charging and real-time response levels. Support for convergence of services (e.g., voice and data). Adaptable federated service mediation facility. Support for a variety of OSS and service value chain.

Federated Accounting Management System Architecture

5. 6. 7. 8. 9.

17

Charge detail aggregation of composite services. Automated inter SP domain accounting and settlement. Interaction with legacy billing systems. Increased demand for guaranteed QoS and related discounting. Rapid service deployment.

record QoS reports

retrieve SLA details

End-User

calculate charges and discounts aggregate usage and send E-IP...

service package

view bill online perform SP settlement send invoice and collect payment

manage customer account balance

IES Customer

resolve customer queries

control resource usage (access)

IRS Provider

IES Provider

Fig. 3. Federated Accounting Management Use Case Model

Up to this point, this paper has presented the business context and key requirement. The following sections present the information model and design of the accounting management system.

4

Federated Accounting Information Model

This section presents the federated accounting information model. Before providing the details of the model, a brief rationale for developing it is given. If accounting management facilities are to be distributed then the SPs must use an information model that: (1) adapts easily to new services; (2) maintains consistency of information during exchange; (3) helps SP in using the information generated by other SPs with little or no effort. Clearly, a common structure, semantic, and syntax of information is crucial in a successful acceptance and functioning of federated accounting management and this is where the federated information model comes into play.

18

4.1

B. Bhushan et al.

Enhancements to IPDR Master Schema

The governing idea behind federated accounting management is to support distribution of service mediation and rating facilities across several SP domains. A note for those readers who are familiar with FAB of TMForum, this support means enabling accounting management at BSS level. This is the level at which rates are applied, charges calculated and bills presented. This is required for the aggregation of charges when a service package is provided to the customer. This package is an agreement that allows the customer to use many services. Several SPs may be involved in providing the package. Clearly SPs have a stake in the bill payments that IESP receives from customer. SPs themselves may want to calculate charges for the services that they contribute towards the service package. IESPís role in this scenario is to aggregate charges and do SP payment settlement. Therefore SPs require means by which they can exchange usage data as well as charges in a single IPDR document. This requirement means that information model must be able to support exchange of usage data as well as charge details. In order to meet this requirement, this paper proposes federated accounting information model. The model is an enhancement on Master IPDR Schema [9] with a CE (charge entry) has been added to the latter. This enhancement enables IESP to the charge aggregation and settlement in a multi-SPs environment. The CE holds information such as the charge that the customer owes to SP, discounts provided by the SP, etc. These details are received by IESP and used during SP settlement and payment. If several SPs send their charge details, all charges are aggregated and user pays a single bill for a service package. The Master Schema that the IPDR organisation has produced does not provide for such requirements because it is outside the scope of the organisation. The federated accounting information model, which is based on Master IDPR Schema, does provide for such requirement. SPs can use CE (charge entry) E-IPDR document (E stands for enhanced) to include charge details and transfer them with IESP. 4.2

Federated Accounting Information Service-Specific Using UML Notations

It must be noted that IPDR information model and all model derived from it are implemented using XML Schema Recommendation. This essentially gives us three schemas, which follow from each other in this order: (1) Master IPDR Schema; (2) Federated Accounting Information Schema; (3) Federated Accounting Servicespecific Schema. Master IPDR Schema: This is the output of IPDR organisation and the basis for all service-specific schemas and federated accounting information schema. This schema has a service-specific part, which can be adapted to any service. SPs can use this schema and adapt its service-specific part to any service for which usage mediation is required. Since federated accounting information schema is based on Master IPDR schema, it is worthwhile to mention it here. For a fuller description, please refer to IPDR specification [9]. Federated Accounting Information Schema: This schema is essentially the Master IPDR Schema, plus the CE element. The CE element is optional and gets populated when SP wishes to calculate its own charges, leaving charge aggregation to IESP.

Federated Accounting Management System Architecture

19

Federated Accounting Service-Specific Schema (see Fig. 4): This schema inherits all the elements of Federated Accounting Information Schema, including CE. It is adapted to provide for and record usage events that are specific to a service. For the development of federated accounting management system, two service-specific schemas have been developed; they are for MediaShop (shown in Fig. 4) and VoIP service. Usage attribute specific to these two services are added to the ones that are inherited from federated accounting information schema. The usage information forms the content of E-IPDR Document.

IPDRRecList

1..n

0..1

IPDRRec id 0..1 startTime info

IPDRRecRef 1 ref

1

IPDRSchema IPDRDoc

choice

IPDRDoc seqNum version startTime 1..n info subscriberId***

IPDR choice id 1..n time seqNum 0..1

1 SS

SCRef ref SC

choice 1

1 id service choice

SC_MediaShop userID subscriberID contextID userApplicationHost

CE 1 customerCharge 1 SERef customerDiscount SE 1 settlementCharge ref UE settlementDiscount chargeTime SE_MediaShop currency serviceID UE_MediaShop serviceTariffID type serviceProviderID serviceProviderHost 1 1 1 1

BaseIPDR refs

0..1 IPDRDoc.End count endTime

1

SSRef ref

0..1

uploadContent resourceName 0..n 0..n contentSize downloadContent timeUploaded resourceName timeDeleted downloadStartTime downloadStopTime contentSize

0..n deleteFolder

0..n createFolder

Fig. 4. Federated Accounting Service-Specific Schema for MediaShop Service

Fig. 4 serves three purposes. Firstly, it illustrates how Master IPDR Schema has been used to implement federated accounting information model. One can see the position and contents of CE element in the Fig. 4. Secondly, it also illustrates how federated accounting information model can be specialised to a particular service or network. It also shows how usage information generated by MediaShop can be modelled. Finally, to best of authorsí knowledge, no UML diagram of IPDR Master Schema exists. Therefore the Fig. 4 allows UML users to see how various XML elements of IPDR Schema fit together to form a complete E-IPDR document. As the reader will see in the next section, federated accounting service-specific schema is an integral part of the system architecture and E-IPDR documents are the information objects communicated at the interfaces such as interdomainAcctMgmt, Rating Bureau Service Interface. All E-IPDR documents are XML instance

20

B. Bhushan et al.

documents. Information objects (mainly usage and charge details) are integrated under a single tree-shape structure. Integration is done at the level of federated accounting management system boundary. Objects and structure are shared by all system interfaces, a shared information schema.

5

System Architecture

This section provides the definitions of components and interfaces of the federated accounting management systems architecture. The architecture is specified using UML (Unified Modelling Language) therefore a mapping of the components and interfaces onto to UML artefacts, namely boundary class, control class and information object is also briefly discussed (see Fig. 5).

OnlineBillingManagem ent Composite E-IPDRDoc

queryDoc

QueryEngineInterface

boundary class

Query Engine

control class

information object

E-IPDRDoc

MediaShop Mediation Adaptor

Federated interdomainAcctMgmt E-IPDR Recorder Mediation Adaptor

Rating Bureau Service Interface

VoIP Mediation A daptor

Rating Bureau Service

Fig. 5. Components of Federated Accounting Management System Architecture

An interface can be considered a boundary class. A component itself can be considered a control class. The term boundary has a strong connotation of something facing outside or an interface. Among other matters, an interface deals with communication protocols. If communication protocol changes, we need to change the interface; we need not change the component that offers the interface. The term control class has a strong connotation of something that receives things and manipulates them without paying much attention to how they will be sent to other components. The things that control objects manipulate are information objects. A component deals with co-ordination task, leaving communication task to its interface. A component can have its own state, change its state independently of other components, and can send external event whenever it wishes to do. If co-ordination task changes we may need to change the component but we can still use the same interface offered by the component.

Federated Accounting Management System Architecture

21

It co-ordination work is not complicated and sizeable enough then we may not need control object at all. A boundary object will be enough to do communication as well co-ordination. 5.1

Interfaces (Boundary Objects)

InterdomainAcctMgmt: This interface resides between IESP and third-party providers and supports usage mediation in a federated SPs environment. This interface is provided by the Federated Mediation Adaptor (FMA) component. Mediation Adaptors (MAs) can use this interface to send E-IPDR documents to FMA. OnlineBillingManagement: This interface resides between IESP and IES Customer. This is the customer entry point to the charging information stored by the IESP. 5.2

Components (Control Objects)

Federated Mediation Adaptor (FMA): This component supports collection and aggregation of E-IPDR documents and generation of a composite E-IPDR document. The documents are aggregated under a single usage session (or parent session), which denotes the beginning and end of service package use. MediaShop Mediation Adaptor (MA): MediaShop MA collects usage data, records the usage events, generates an E-IPDR document, and sends it to FMA. VoIP Mediation Adaptor (MA): The main role of VoIP MA is to collect usage data, record the usage events, generate an E-IPDR document, and send it to FMA. QueryEngine: It ensures the availability of all recorded IPDR information. E-IPDR Recorder: The IPDR Recorder takes upon the role of IPDR documents recorder/transmitter. It implements method specifies in the IPDR ìProtocol Primitives and Parametersî specification (Push, Pull, Subscribe, etc). Rating Bureau Service (RBS): RBS applies tariffs, calculates discounts and calculates charges for individual services that make up a composite service, for example, OCS (i.e., Online Collaboration Service mentioned in Sect. 3). Then all individual charges are aggregated into a single composite service charge. Finally, RBS generates and stores a rated/discounted E-IPDR document. 5.3

Information Objects (E-IPDR Documents)

Composite E-IPDR Doc: This is a list of E-IPDR documents, each of which is an instance document of federated accounting service-specific schema. Its main role is to carry usage information detailing usage of a composite service, as opposed to usage information for a singleton or simple service. QueryDoc: An XML instance document, containing end-userís query. E-IPDR Doc: This is single E-IPDR document, based on federated accounting service-specific schema. Its main role is to carry usage information for a singleton service. This information is details of usage events that are generated when the enduser uses MediaShop or VoIP services.

22

6

B. Bhushan et al.

System Integration and Technology Architecture

This section discusses how a prototype federated accounting management system was implemented. Fig. 6 shows the technologies that have been used to implement and integrate components of the prototype.

OnlineBillingManagement

JSP/Servlets, T

E-IPDRDoc Composite E-IPDRDoc

MediaShop MediationAdaptor

Java/CORBAIIO P/ Oracle

queryDoc

QueryEngineInterface QueryEngine

EJB/J2EE S

E-IPDR Recorder

Java/WSDL, SOAP

Federated Mediation Adaptor

Interdomain AcctMgmt

Java/CORBA IIOP/O l Rating Bureau Service Interface

VoIP Mediation Adaptor

C++/SuSE 0

C#/C++ COM SOAP WSDL Rating Bureau Service

Fig. 6. Federated Accounting Technology Architecture

The main components (see Sect. 5.2) of the prototype were implemented using C++, Java, and C# and used IIOP, RPC COM/CORBA-SOAP for communication with one another. Interfaces were specified using CORBA IDL and WSDL. Information objects (or E-IPDR and Composite E-IPDR documents) exchanged between components are XML instance document. They are modelled on federated accounting service-specific schema, which is specified using W3C XML Schema, Structure and Data Types. Oracle is used as storage for E-IPDR documents. The technology described here has yielded many benefits. Java and EJB make architecture components highly portable. XML/WSDL allows service providers to deploy accounting components as Web-services and seamlessly exchange accounting information. Oracle enables direct mapping of IPDR structure (XML documents) onto database tables. SOAP binds IPDR documents and HTTP tightly, making document exchange them over the internet easy. OSP (Open Service Platform) is a service subscription and access management middleware, which is being developed by Fraunhofer-FOKUS. It is based on TSAS (Telecommunications Service Access and Subscription). A number of lessons were learnt from the development and evaluation of the federated accounting system. Some important lessons are discussed here. An interface at the boundary between two SP may possess more functionality then assigned during design phase. For example, interdomainAcctMgmt interface may also

Federated Accounting Management System Architecture

23

monitoring and recording network QoS deterioration. Interfaces implementations are kept simple because of the more complex structures (defined in XML) that are passed. Standardised information model (i.e., IPDR Master Schema) proved to be useful in developing federated accounting service-specific schema. It also reconciles two differing sets of concerns (information-wise), service mediation and charge details that lead to a common goal, i.e., Billing. There are trade-offs when using XML schemas, which are more complicated than DTDs (Document Type Definitions) but are richer and useful in expressing a vocabulary of federated accounting management business process.

7

Conclusions and Future Work

This paper has presented a requirement analysis model and architecture of federated accounting management system. The system presented has been developed and tested and evaluated in two trials. In order to ensure that components and interfaces provided the required functionality, test case result were mapped onto requirement captured in the requirement analysis phase. Mapping resulted in a set of key requirements that were important for the development of federated accounting management system. They are listed in Sect. 3. One of the prime aims of the research and development was to validate the specification being developed in standardization community and feed the result into standardization process. It is fair to say that authors have successfully achieved this aim. Results obtained have been provided as contributions to a number of international organizations. A significant proportion of design specification has been provided as input to Telecom Domain Task Force of OMG, which is also developing a specification of federated charging and rating facility. An initial business model and use cases have been provided to AAA WG of IRTF (Internet Research Task Force). A contribution containing intermediate business model and informational model has been provided to TMForum through a TM World conference. There are further plans to provide an enhanced IPDR Master Schema as an input to IPDR organization. Several areas in which future work can be carried out have identified. Authors consider the use of XML and Webservice technologies (e.g., XSLT, WSDL, SOAP) for component and interface definitions an important area for future work. Therefore further research into Web service technologies will be carried out. Use of XMI and XSLT in transforming technology-neutral system architecture to a technology-specific system is essential if federated accounting management is to appeal to E-commerce. Another area, which is gaining acceptance, is ebXML. It will be used for the definition of federated accounting management business process. Further development will also be carried out into a generic approach to aggregated services mediation. In this generic approach, functionality will be added to FMA component, thus making them adaptable to several different types of communication protocol and technologies. Components and interfaces will be further enhanced to support QoS and a guaranteed delivery of E-IPDR documents. Last but not the least, the authors would like to thank the members of PLATIN group (Fraunhofer-FOKUS Germany) and TSSG group (WIT Ireland) for their

24

B. Bhushan et al.

constructive comments. This research work was carried out under the auspices of EUsponsored IST project FORM (IST-1999-10357). More information on FORM can be obtained from www.ist-form.org or from the authors.

References 1. Varshney, et al, ìVoice Over IPî, Communications of ACM, Jan 2002,Vol 45,No 1. 2. Bartholome, Lloyd, et al, ìA Practical Approach For Implementing E-Commerce Programs in Business Schoolî, Communications of ACM, January 2002, Vol. 45, No. 1. 3. Howard-Healy, ìBilling the future of contentî, Billing, Issue 7, September/October 2000. 4. Browne, ìHave you got money to burn?î, Telecommunications International, Aug 2001. 5. Whitworth, ìThe Unknown World of 3G Tariffsî Telecommunications International, Feb 2001. 6. Dobson ìThe SLA Battlegroundî, Telecommunications International, March 2002. 7. Telecom Operations Map, Approved Version 2.1, TeleManagementForum, March 2000. 8. TMForum GB921: ìeTOM: The Business Process Framework For The Information and Communication Services Industryî, Members Evaluation Version 2.0, October 2001. 9. Network Data Management ñ Usage (NDM-U): For IP-based Service, Version 2.6, October 2001, IPDR Organisation. 10. RFC 2924: Accounting Attributes and Record Format, Category: Informational, Authors: N. Brownlee, A. Blount, September 2000. 11. RFC 3127: Authentication, Authorization, and Accounting: Protocol Evaluation, Category: Informational, Authors: Mitton, Johns, Barkley, Nelson, Patil, June 2001. 12. Bharat Bhushan, et al, ìFederated Accounting: Service Charging and Billing in a Businessto-Business Environmentî, IM2001: 2001 IEEE/IFIP International Symposium on Integrated Network Management Proceedings, Seattle, May 2001. 13. Huston, Geoff, ìInterconnection, Peering and Settlements: Part-Iî, The Internet Protocol Journal, Volume 2, Number 1, March 1999.

Policy-Based Quality of Service and Security Management for Multimedia Services on IP Networks in the RTIPA* Project ValÈrie Gay1, Sandrine Duflos1, Brigitte Kervella1, Gladys Diaz2, and Eric Horlait1 1

LIP 6 : Laboratoire d'Informatique de Paris 6, 8, Rue du capitaine Scott, 75015 Paris, France {Valerie.Gay, Sandrine.Duflos, Brigitte.Kervella, Eric.Horlait}@lip6.fr 2

L2TI ñ Institut GalilÈe ñ UniversitÈ Paris 13, 99 Av. J. B. ClÈment, 93430 Villetaneuse - France [email protected]

Abstract. This paper summarizes the research work that has been conducted in the context of the RTIPA project on policy-based QoS (Quality of Service) and security management for distributed multimedia services. It presents an architecture allowing the derivation of policies from the service level down to the network level. It is a step towards an end-to-end QoS and security management for distributed multimedia services running on the new generation of IP networks.

1

Introduction

Enabling end-to-end QoS and security management of distributed services is a research topic that has been going on for some years now. Services are getting adaptive and QoS-aware and the network is now able to deal with different levels of QoS and security. However, there is still a gap between the services and the network, and achieving end-to-end QoS and security for distributed applications is still bringing some challenges to the middleware and network research communities. In the framework of the RTIPA project [1], our role was to reduce that gap and to make a step towards end-to-end QoS and Security management for distributed multimedia services running on the new generation of IP networks. In the project we dealt with policy-based network management therefore we naturally focused our research on a policy-based solution. The major part of current standardization works on policy-based management is done at the Internet Engineering Task Force (IETF) and are positioned at the network level. The IETF proposes a centralized management [2]. This model is composed of several entities: the Policy Repository (PR) where policies are stored, the Policy Decision Point (PDP) The original version of this chapter was revised: The copyright line was incorrect. This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 *

RTIPA: Real-Time Internet Platform Architectures, ITEA (Information Technology for European Advancement) project

K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 25-35, 2002. © IFIP International Federation for Information Processing 2002

26

V. Gay et al.

that decides the policies to apply, and Policy Enforcement Points (PEP) that enforce the policies. At the network level, IETF also defined an information model: The Policy Core Information Model (PCIM) [3] [4]. It has also defined extension for QoS management: The QoS Policy Information Model (QPIM) and the QoS Device Data path Information Model (QDDIM) [5] [6]. For security management, the IPsec Configuration Policy Model (ICPM) [7] describes the information for policy management. Other research works propose solutions to specify, manage and enforce policies at network or middleware level. Examples are found in [8] that describes a solution for IPsec policies, [9] that bases its solution on the Corba Security framework [10] and [11] that proposes an object-oriented solution to manage and enforce policies expressed in a language called Ponder [12]. Our work presents an architecture for end-to-end QoS and security management of distributed multimedia services. It focuses on the top-down aspects of the policy management and its relationships with security and QoS managers. It also describes how it could be implemented on an Ipv6 network that takes into account the service level policies thanks to a refinement process. This paper describes in Section 2 an architecture for policy-based management of QoS and security for multimedia services. Section 3 describes the policy refinement process. Section 4 presents how the policies are supported at the network level on the RTIPA platform. Section 5 concludes on open issues and perspectives

2

An Architecture for Policy-Based QoS and Security Management

Our architecture is presented in Fig. 1. It is separated into three abstraction levels: service, middleware and network. Fig.1 only highlights the elements and interactions we are discussing in this paper. Other managers that do not appear in the figure include the billing, mobility and Service Level Agreement (SLA) managers. At the service level, QoS, security, and policy rules are expressed in high-level terms understandable by the end-users. A QoS contract specifies the service requirements and offerings in terms of QoS and security. Its granularity can vary from 'per service' to 'per flow type'. The service level policy rules can express general administration rules (such as routing policy rules in the case of network management) in high-level terms. The policy manager maps them onto middleware policy rules. The service level QoS, policy and security requirements and offerings are translated into middleware level policy rules (using the middleware QoS, security and policy managers) to be integrated into the new set of middleware policy rules. The middleware policy manager checks the policy rules and resolves the conflicts. Feedback can be provided to the application level and can lead, for example, to the renegotiation of QoS or price or to the adaptation of the content. The QoS and security managers check if the different QoS and security requirements specified at the service level can be met by the infrastructure and provide feedback to the application on the status of the infrastructure with respect of the application QoS and security. The resource manager [13] has a view of the available resources. It enables the QoS and security managers to know if their QoS

Policy-Based Quality of Service and Security Management for Multimedia Services

27

and security can be fulfilled. Especially for real-time service it is important to have strict resource management policies so the real-time QoS requirements are valid during the lifetime of the service.

Service management

QoS contracts (including Security)

Policy rules

Services

Requirements Negotiation Feedback

Middleware management

QoS manager

Policy manager Translation Refinement Updating Conflict resolution

Security manager

Policy rules

Resource manager Requirements Negotiation Feedback QoS manager

Network management

Security manager

Policy manager Refinement, Updating Conflict resolution Policy Dissemination rules Enforcement

Fig. 1. Policy-Based QoS and Security Management Architecture The middleware level policy rules are then translated into network-level policy rules. The network policy manager adapts the rules to reflect the requirements it gets from the other network managers (such as a network QoS manager). It is disseminating the rules and enforcing the policies. It resolves the conflicts between the policy rules and provides feedback to the middleware managers. QoS management and mapping have been discussed in our previous work [14]. In that paper, the content of the QoS contracts are detailed, they express the QoS

28

V. Gay et al.

requirements and obligations of the distributed services. The QoS manager [13] monitors the network QoS and can initiate operations like QoS re-negotiation in case of QoS violation. It checks if the environment can and will fulfill the contracts. It requests the resource manager for certain resources (e.g. buffer, CPU-time, Audiodevice) that can be allocated to a certain service. There is a need of cooperation of QoS managers in different domains. The goal of the security manager is to provide the necessary security that is required by the multimedia service. The security manager analyses and translates these requirements, expressed in the QoS contract at service level, in security parameters depicting the security aspects to provide. For example, in the case of a Video-on-Demand (VoD) service, a minimum of security is necessary to protect the audio/video streams against a fraudulent listening. In such a case confidentiality is necessary. The resource manager is checked with to see if a security protocol can set this security aspect. If yes, it sends the parameters to the policy manager that integrates them in the middleware level policy rule. In order to produce the network level policy rules, the network level security manager must be able to provide the security protocols and cryptographic algorithms available to set the security required. In our VoD example, the confidentiality parameter becomes ESP and (3DES, DES) ESP being the IPsec protocol providing confidentiality and 3DES and DES a proposition of associated cryptographic algorithms. The security and policy managers interact a last time for the protocol configuration, in our example IPsec, when the network policy rule is disseminated. In case the policy managers cannot resolve the conflicts, the security managers must be able to provide an alternative solution (e.g.: security parameters renegotiation at middleware level or proposition of others protocols and cryptographic algorithms at network level) to produce a new policy and to secure the service. If it is not possible the service should be turned off. The managers shown in Fig. 1 would typically be located on every node and cooperate among each other to manage the distributed multimedia service. We are not describing their internal mechanisms in this paper. Policy rules [15] evolve through a life cycle (specification, refinement, static and dynamic conflict resolution, dissemination, enforcement, updating (includes disabling, withdrawing and deleting). Actions related to this life cycle are shown in Italics in Fig. 1. In the following section, we focus on the refinement process.

3

The Policy Refinement Process

This section describes the policy translation and refinement process from the middleware to the network levels. Our work is based on the TMN (Telecommunications Management Network) architecture defined by the ITU-TS [16] and the research work of [17]. The different managers presented at middleware level derive from the service level the rules to be generated and included in the middleware-level policy rules. A possible template for the middleware level policy is described in Table 1 where is the network configuration type (e.g.: point to point, multicast, star, fully meshed.), is the

Policy-Based Quality of Service and Security Management for Multimedia Services

29

requested transport service (e.g., ATM, SDH/PDH, IP), is the requested application (real player, Netscape, video on demand), and are the identified users who will participate to the application. is the requested QoS level (e.g.: premium, gold, silver, bronze). is the requested security configuration including confidentiality, authenticity (integrity and authentication of packet sender), mutual authentication, non-repudiation, tunnel mode and anti-replay. The different parameters values are checked by the middleware managers described in the previous section (e.g.: are the recorded members known by the service? is the requested security configuration correct?) Once the request is validated, it can be translated into a middleware-level policy rulerepresented here under the form "If condition Then provide configuration" that will support the implementation of the QoS and security requested. Table 1. Middleware level policy IF SourceUser|users = AND DestinationUser = AND ApplicationType = THEN PROVIDE of for

with Guarantee

Quality and Security The middleware level policies are only understandable at that level. They use middleware-level parameters such as a specific security configuration (such as confidentiality) or the users' name. Some parameters do not have to be managed by the network because they are dealt with at a higher level but the rest of the parameters have to be converted in network level parameters. In the RTIPA project, our network is based on IPv6 therefore we are using IP Diffserv and IPsec [18] [19] [20] [21] at network level. A simplified translation table is presented in Table 2 Table 2. Simplified Translation from Middleware to Network-Level Parameters User Application

Middleware level parameters (e.g.: End-User1)

(e.g.:VoD)

Configuration and and transport (e.g. point to point, IP) QoS (e.g.: gold) Security

(e.g.: confidentiality)

Network level parameters (e.g.: 2.2.2.2)

(e.g.: 8000, uni, unicast, 2.2.0.0)) Identical (e.g.: AF11) , , , , (e.g.: ESP, (3DES, DES), NULL, IPsec_transport, NULL)

30

V. Gay et al.

In Table 2 and Table 3, corresponds to the users IP addresses, is the port used, is the QoS direction (unidirectional or bi-directional) the is the connection type (e.g. unicast, broadcast, multicast) and is the desired Per-Hop behavior. is the IPsec security protocol used (AH or ESP), is a list of algorithms used for confidentiality (e.g.: {private, key cryptography} or NULL), is a list of algorithms used for authenticity (e.g.: {digest function, public key cryptography} or signature cryptography or NULL), is the security mode (tunnel or transport) and is set to True or False to indicate that it can or cannot be replayed. For instance, the security parameters, expressed at the middleware level, will be replaced by the IPsec security protocol (Sec-Prot), a list of algorithms used for confidentiality (C-Algo), a list of algorithms used for authenticity (A-Algo), the security mode (tunnel or transport) used for the creation of secure tunnel (e.g. VPN) and the anti-replay protection (No-Replay). Furthermore, a user will be replaced by its IP address, the QoS type by the corresponding DiffServ PHB (Per Hop Behavior) type, etc. This modification of parameters allows the co-ordination of the policy implementation in the whole network. The new policy rule template is shown in Table 3. Table 3. Network-Level Policy IF SourceIPaddress|UserIPaddresses = and SourcePortNo|UserportNo=and DestinationIPAddress = and DestinationPortNo = THEN CONNECT with and from|among at to at with and with and and and A last step is necessary to disseminate and enforce the policies rules in the different network entities. Only the rules and parameters of concern for the entity where the policy will be enforced are disseminated. Table 4. Network-Level Policy for Dissemination to the Network Elements IF SourceIPaddress|UserIPaddresses = and SourcePortNo|UserportNo = and DestinationIPAddress = and DestinationPortNo = THEN SET at With and with and and and For example for a videoconferencing service, when two users from distinct domains (managed by distinct entities) communicate, each entity receives the IP address of the other user. This policy rule allows the definition and the establishment of security associations between the communicating entities to ensure an end-to-end security. Table 4 presents the policy that will be implemented in the network elements. The keyword SET makes it clear that the objective is no longer to provide

Policy-Based Quality of Service and Security Management for Multimedia Services

31

something but to set something to certain values. The represents where the policy must be enforced. It generally represents an edge router on which a set of is connected. This is a first step towards automatic translation. Ideally, the refinement process could be automated through the use of a translation table as depicted in Table 2 but the translation process is more complex than that. In addition there is a need to resolve conflicts and to deal with the fact that some resources might not be available.

4

Network Level Prototyping

In the RTIPA project, we worked on different aspects of the architecture. One of our task was to study languages mapping between applications using SMIL [22][23] and the middleware level. However, the network level was the only really stable basis for the implementation. Standards on network policy management exist and the choice of technology had been made [24]. It is therefore on that level that we chose to demonstrate some aspects of our architecture. Our objective was to integrate QoS and security policies using Traffic Designer, a network-level tool developed by Qosmos [25]. This section describes how Traffic Designer is used to support policy-based QoS and security management at the edge routers level. The architecture of the Traffic Designer (TD) and its interactions with the different managers of our architecture are depicted in Fig. 2. Traffic designer is an integrated tool able to provide actions on the flows (classification, prioritization, throughput guarantee, forbiddance, tagging of packets, counting, analyzing and spying) based on a set of rules. This set of (policy) rules integrates the policy rules that have been derived as described in section 3 (Table 4). Traffic designer is composed of a classifier (TD classifier), a filter (TD filter) and five action modules (TD QoS, TD Firewall, TD cryptography, TD Diffserv and TD Monitor and its related TD log database). The TD classifier takes as input packets or packet flows and gives as an output the highest level of protocol the packets belongs to. This allows us to provide rules on those particular applications. The TD filter applies the rules (including our policy rules) and redirects the packets to the action modules if appropriate. Examples of rules are: •IF (udp_dport=1024..65535) AND (tcp_dport=1024..65535) Then goto Firewall (rule based on the header fields)

•IF (http_mime=" image/gif ") AND (smtp_attach=" application/pdf") Then provide 20 Kbps (rule based on the application properties)

•IF (flow_content=*topsecret*) then goto Firewall (rule based on the packet content (not advised))

•IF (client to server) then...(rule based on the direction of a TCP connection). The TD QoS module is able to shape the traffic, to apply queuing mechanisms and to recover the unused bandwidth. It gives the possibility to guarantee a minimum or a maximum throughput, to tune the maximum waiting delay for the packets, to provide

32

V. Gay et al.

priorities within a queue (e.g. LIFO). The packets can then be tagged depending on applicative criteria. It enable us, in particular, to guarantee bandwidth to mission critical applications and establish classes of traffic depending on the protocol used and the user involved. Requirements Negotiation Feedback

Policy Dissemination

Node Management (Edge router) Policy rules

Policy manager Conflict resolution

QoS manager

Security manager TD QoS

TD Classifier

TD Log TD Classifier Filter (PEP)

TD Cryptography

TD Flow Monitor Packets TD Firewall Firewall

TD DiffServ

Traffic Designer Fig. 2. Use of Traffic Designer at the Edge Routers The TD firewall (security) can drop packets. It provides us with functionalities to protect the application against hackers, to filter per protocol or per application URL and keywords, to position alarms on certain event. This security aspect is interesting but not sufficient for our security needs. To enforce our security policy rules we need another action module, the TD cryptography module that is currently under development. This new module will provide a user-transparent cryptography (multiple key system, choice of the key depending on the application or of the user). When available, this feature will be

Policy-Based Quality of Service and Security Management for Multimedia Services

33

studied in order to know if it can be used for the cryptographic algorithm management. The TD Diffserv module adapts the flow to DiffServ. The TD Monitor provides information and statistics in its TD log database (number of packets/bytes, instant/average/peak throughput, percentage of bandwidth used, zoom on a particular traffic type and spying, discovery based on applicative criteria). The syntax used to query the database is a simplified subset of SQL. Some examples are: • SELECT (tcp_sport, ip_saddr) FILTER (tcp_proto=http) AND NOT (ip_saddr=" webserver ") Result: o 8080 192.168.2.5 79pkt, 10,5Kbps, 4342octets o 7979 192.168.2.8 4 pkt, 2,1 Kbps, 324 octets

• SELECT (smtp_sender) FILTER (smtp_subject=" *big brother*") OR (smtp_subject=" *loft story*") Result: o [email protected] 79 pkts, 0 Kbps, 4342 octets o [email protected] 458 pkts, 2,1 Kbps, 33424 octets o [email protected] 2303 pkts, 0 Kbps, 23423 octets

The policy manager acts on the TD filter to enforce the policy rules. The QoS and security managers of the edge router interact with the TD Monitor to describe the information to track. They can then query the TD Log database to get the information they need and if necessary they can provide a feedback to the network, middleware and service levels.

5

Conclusion: Open Issues and Perspectives

This paper has presented our research work done in the context of the RTIPA project to manage QoS and security for distributed multimedia services running on IP networks. Some open issues still remain. The services have a range of QoS requirements and offerings well identified. Various protocols and mechanisms exist to support QoS in distinct architectures. But the problem is how can we mix and match their capabilities to provide a complete QoS and security management? For example, at the network level, some incompatibilities appear between the IPsec and Diffserv protocols. The refinement of high-level policies (service level policies in our architecture) is another issue: it is difficult to automate. It is likely to remain partially a manual process since it requires a large amount of knowledge. Also, what happens when the different rules cannot be validated (conflict between the rules...) or when the network cannot support the low-level policy? What sort of feedback should be provided to the user and will there be renegotiations of the QoS and security requirements and offerings? If it is the case, a relationship binding the different policy abstraction levels, as in [26] with the parent/child relationship is essential. In this article, we have depicted a simplified version of QoS and security policy rules but a problem is observed during inter-domain communications. The domains

34

V. Gay et al.

do not have necessarily the same way to represent policy, QoS and security. Several policy specification languages exist such as [27],[28],[29],[30],[31]. They provide flexibility and permit to express policies through different approaches (logical, graphical, etc.). Ponder [12] is the most suitable language for our solution. It integrates QoS and security policies, it is mature and an implementation is currently available. Another candidate for the security aspect would have been the Security Policy Specification Language (SPSL). There is a need for a global agreement on the policy language, and this not only at the network level, to be able to implement an end-to-end policy-based QoS and security management. Our objective is to continue this research work and to extend this architecture to SLA and mobility management. The work on security is further detailed in another paper [32].

Acknowledgements The authors would like to thank the members of the RTIPA project, Qosmos and the APS group of the university of Twente for their valuable comments on this work.

References 1. RTIPA http://www.extra.research.philips.com/euprojects/rtipa/ or http://www.iteaoffice.org/projects/facts_sheets/rtipa_fact_sheet.htm 2. R. Yavatkar et al., A Framework for Policy-Based Admission Control, RFC 2753, January 2000 3. B. Moore et al., Policy Core Information Model - Version 1 Specification, RFC 3060, February 2001. http://www.ietf.org/rfc/rfc3060.txt. 4. B. Moore et al., Policy Core Information Model Extensions V.7, http://www.ietf.org/internet-drafts/draft-ietf-policy-pcim-ext-07.txt, March. 2002. 5. Y. Snir et al., Policy QoS Information Model, Policy framework Working Group, Internet Draft, http://www.ietf.org/internet-drafts/draft-ietf-policy-qos-info-model-04.txt, November 2001. 6. B. Moore et al., Information Model for Describing Network Device QoS Datapath Mechanisms, Policy framework Working Group, Internet Draft, http://www.ietf.org/internet-drafts/draft-ietf-policy-qos-device-info-model-08.txt, May 2002. 7. J. Jason et al., IPsec Configuration Policy Model, Internet Protocol Security Policy Working Group, Internet Draft, February 2002. 8. J. Zao et al, Domain Based Internet Security Policy Management, Proceedings of DARPA Information Survivability Conference and EXposition 2000 (DISCEX '00), January 2000. 9. R. Chandramouli, Implementation of Multiple Access Control Policies within a CORBASEC Framework, 22th National Information Systems Security Conference NISSCí1999. 10. Corba "Security Service" Specification v 1.7, OMG, March 2001 (ftp://ftp.omg.org/pub/docs/formal/01-03-08.pdf) 11. N. Dulay et al., A Policy Deployment Model for the Ponder Language, 7th IFIP/IEEE International Symposium on Integrated Network Management (IM'2001), Seattle, USA May 2001.

Policy-Based Quality of Service and Security Management for Multimedia Services

35

12. N. Damianou et al., Ponder: Language for Specifying Security and Management Policies for Distributed Systems, The Language Specification Version 2.3, October 2000. http://www-dse.doc.ic.ac.uk/Research/policies/ponder/PonderSpec.pdf 13. P. Leydekkers et al., A Computational and Engineering View on Open Distributed Realtime Multimedia Exchange, In Lecture Notes in Computer Science - Number 1018 Springer Verlag - ISBN: 3-540-60647-5 - Proceedings of Network and Operating System Support for Digital Audio and Video (NOSSDAV'95), Durham, New Hampshire, USA, April 1995. 14. P. Leydekkers and V.C.J. Gay, ODP View on Quality of Service for Open Distributed Multimedia Environments, In 'Quality of Service - Description, Modelling and Management', A. Vogel and J. de Meer editors. Proceedings of the 4th International IFIP Workshop on QoS, Paris, March 1996. 15. M. D. J. Cox and R. G. Davison, Concepts, Activities and Issues of Policy-Based Communications Management. BT Technology, Vol 17 No 3 July 1999. 16. Jajodia S. et al., Principles for a Telecommunications Management Network, ITU-T M.3010, February 2000. 17. H.J. Tjaja, SLA Enforced by Policy, masters thesis, KPN/Twente University internal report, 2001. 18. K. Nichols et al.., Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers, RFC 2474, December 1998. 19. S. Kent and R. Atkinson, IP Authentication Header, RFC 2402, November 1998. 20. S. Kent and R. Atkinson, IP Encapsulating Security Payload (ESP), RFC 2406, November 1998. 21. S. Kent and R. Atkinson, Security Architecture for the Internet Protocol, RFC 2401, November 1998. 22. J.van Ossenbruggen et al., Towards Second and Third Generation Web-Based Multimedia. In: The Tenth International World Wide Web Conference, May 2001, Hong Kong 23. SMIL 2.0 documentation http://www.w3.org/TR/2000/WD-smil20-20000921/ 24. M. Riguidel, Introduction to Policy Management, RTIPA Technical Report, 2000. 25. www.Qosmos.net 26. D. Marriott, PhD Thesis, Policy Service for Distributed Systems, Department of Computing, Imperial College, London, 1997. 27. N. Damianou, A Policy Framework for Management of Distributed Systems, PhD Thesis, Imperial College of London, February 2002. 28. T. Koch et al., Policy Definition Language for Automated Management of Distributed Systems, Proceedings of the Second International Workshop on System Management, IEEE Computer Society Press, 1996. 29. J. Hoagland, Specifying and Implementing Security Policies Using LaSCO, the Language for Security Constraints on Objects, PhD Thesis University of California, Davis Department of Computer Science, 2000. 30. M. Hitchens et al., Tower: A Language for Role Based Access Control, Policy 2001, LNCS 1995, pp.88-106, 2001. 31. Jajodia S. et al, A Logical Language for Expressing Authorizations. In Proceedings of the IEEE Symposium on Security and Privacy. Oakland, CA, USA: IEEE Press, 1997. p. 31-42. 32. S. Duflos et al., An Architecture for End-to-End Policy-Based Security Management, Submitted to the eighth IFIP/IEEE International Symposium on Integrated Network Management (IM 2003), March 2003.

Formal Modeling of Service Session Management M. van Le, B.J.F. van Beijnum, and B.L. de Goede Department of Computer Science, University of Twente, The Netherlands. {le,beijnum,goede}@cs.utwente.nl

Abstract. This paper proposes a concept to apply modeling tools to MultiProvider Telematics Service Management. The service architecture is based on the framework called ìOpen Service Componentsî which serves as building blocks to compose end-to-end telematics services in terms of service components offered by different service providers. Our work presented in this paper contributes to the abstract way of modeling end-to-end Service Management using Architectural Description Language and an underlying Formal Description Language.

1 Introduction Today telematics services are becoming increasingly complex. For instance, new services are being introduced by integrating existing services. Also, in the provisioning of a service session, many different 'service components' have to cooperate in order to establish the end-to-end service session to users. Moreover, instead of having the ability to manage individual service components, end-to-end service provisioning will require the management of the cooperation between these components in a service session. One of the complicating aspects is that in the provisioning of the end-to-end service, multiple providers are involved. Hence, inter-domain management is required. With the advent of user mobility and in-session seamless roaming capabilities new challenges in end-to-end service management are faced [1, 3, 4]. To deal with the complexity of service management, and to arrive at new management solution we envisage an increasing need for the application of existing (formal) modeling and analysis methods. In this paper we present the application of such a method and tools for a (simple) multimedia service provisioning scenario. In future work, specific management issues will be addressed for this and other more complex scenarios, specifically we will address service accounting issues. The modeling and analysis method selected for this paper consists of the architecture description language (ADL) and a formal specification language (FSL). The ADL Darwin is used to express the service in terms of 'components' and the relations between them. To specify the behavior of each component, the formal specification language FSP (Finite State Processes) is being used. The Service Session Management model of the scenario considered basically specifies the coordination of activation, modification and termination needed for the end-to-end service session. The remainder of this paper is organized as follows: The original version of this chapter was revised: The copyright line was incorrect. This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 36-48, 2002. © IFIP International Federation for Information Processing 2002

Formal Modeling of Service Session Management

37

Section 2 discusses a scenario of multi-provider service architecture based on ìOpen Service Components Architecture Frameworkî. Section 3 describes Service Session Scenario and associated Management Model. Section 4 presents the modeling method applied. Section 5 presents the model and some preliminary analysis results of the service session scenario. Finally, section 6 provides with conclusions and outlooks on future work.

2 A Service Architecture The idea of Open Service Component Architecture Framework (OSCAF) is based on operational relevance. The framework is based on the observation that end-to-end services are built from 'service components' [2]. The services offered are mandated by a Service Level Specification (SLS), which is an agreed technical specification of the offered service. This technical specification determines the offered functionality and quality of service characteristics (the SLS is possibly parameterized, thus controllable to some extent by users). The means to offer a service is through the use of service components. In OSCAF four different classes of service components are distinguished, these are: • Connectivity Service Components - Connectivity service components provide the distant exchange of information between geographically dispersed points. Thus, CSCs cover a very broad range of components, varying from bit-pipes, data-links, host-to-host communication, reliable inter-process communication channels, and the exchange of application messages. • Network Service Components - Network service components offer functionality additional to connectivity service components. Concrete examples of network service components are time dependent routing, number translation, closed user groups, DNS, authorization and authentication. It is possible that different actors provide connectivity service components and network service components. • Enabling Service Components - The execution of Application Service components is being supported by Enabling Service components that provide features related to selection, visibility, accountability and other facilities like adapting shape and language to usersí needs. • Application Service Components - Application service components maintain, generate and provide information or functional capabilities intended for users. Examples are Video, MP3, Web-pages etc. However, the range is much wider. For instance: e-mail, relational database, ticket reservation, virtual workspace and the like. The open character of a decomposition as illustrated in Figure 1 is prerogative for seamless service interworking with other Service Providers. An integral approach of Telematics Service Component Management is to arrive at optimal cooperation of the mentioned classes of services.

38

M. van Le, B.J.F. van Beijnum, and B.L. de Goede

Integrated Man agem ent

do

m

ain

s

Network Service Components

Users & Service Subscribers

Enabling Service Components

Service Components

Application Service Components

of

Connectivity Service Components

Fig. 1. Open Service Component Architecture Framework 2.1 Service Session Life Cycle Three service life cycles can be identified in [2] in the context of operational lifetime, namely: Implementation Life Cycle, Subscription Life Cycle and Service Session Life Cycle. The implementation cycle involves: formulation of business plan, analyzing service characteristics, planning network resources, testing software, installing the service, etc. The subscription cycle is concerned with: contracts toward new service users, providing service users with relevant usage instruction, exchange of cost related data about service users with other domain, etc. The service session cycle involves the actual usage of services. The service session life cycle consists of the following phases: • Service Session Invocation - The service user invokes the Service Session Management by means of a service request. The requested service may be a composition different service components. • Service Session Negotiation - This step enables the negotiation of appropriate service components to establish an end-to-end service session. The parameters of negotiation can be specified in a SLA (QoS, Price, etc.). The Service Session Management makes a selection of service component(s) to build up an end-to-end service session. • Service Session Creation - Since the service session management is assumed to be capable to negotiate, select and drop service components offered by the negotiating service providers, this step confirms the participation of the selected service providers in the provisioning of a service session. • Service Session Provision - Upon the reception of service component confirmation, service instances can be set-up and provisioned by the service providers. Parallel to service provisioning, service accounting can be started to account for the delivered service components. Further, service monitoring is essential to provide feedback to service providers about the delivered QoS, this information can have effect on the adjustment of the price that the user has to pay. • Service Session Termination - Service session termination is the releasing of a session. Termination may be initiated by the service user or by a service provider.

Formal Modeling of Service Session Management

39

3 A Service Scenario 3.1 Content Retrieval Service The process of service component delivery from a service provider to a service user may be divided into two aspects: service usage and service provisioning. Seen from the service usage, two important management information exchanges are generated: the service request from the user to the service provider and the service termination. Seen from the service the service provisioning, service management information must be interchanged between service providers to arrive at end-to-end service. We assume the participation of following actors: • Mobile Network Operator (MNO) • Application Service Broker (ASB) • Content Providers (CPs) • Connectivity Service Broker (CSB) • Internet Service Provider (ISP) 3.2 Content Retrieval Service Management Previous definition of Service Session Life Cycle raises a number of research questions concerning multi-provider service management. A few are listed below: • How to define clearly the service management boundary within which a service component is being delivered? • What management functions are required at each domain in order to cooperate with other domains? • What kind of information needs to be exchanged at each step during the session life cycle? • How can the coordination of local management functions be organized in an endto-end service management context? In principle, two management models from the two extreme ends of multi-provider service management spectrum are conceivable: cascade and hierarchical service management. Cascade service management is a management structure where userís service invocation is (re)directed to a content provider and the content provider figures out how to provision the requested service without further interaction with other involved actors. A good example of cascade service management is the current ìbest effortî Video-on-Demand (VoD) service through the Internet. In most of the cases, the content provider injects an amount of packets into the network and these have to find their way to the next hop until they reach the end terminal. Cascade service management does not concern about the dependency of service components as a whole and the end-to-end service session management. Hierarchical service management is a management structure where the handling of service invocation is processed and monitored. Contrary to the former, hierarchical service management has an end-to-end significance concerning service provisioning and the dependency of service component management functions. Other management

40

M. van Le, B.J.F. van Beijnum, and B.L. de Goede

structures are certainly possible. However, for a comprehensive presentation of our modeling approach, a hierarchical management structure will be discussed further in detail. The service management topology illustrated in Figure 2 reflects the management architecture to be modeled in section 5. Three classes of management function are distinguished: 1. Component Management Function (CMF) represented by a dot: components having this class of management function can only act as managed entity. 2. Sub-Integral Service Management Function (sub-ISMF) represented by a square with an internal dot: components having this class of management function can act as managed entity as well as managing entity. 3. Integral Service Management Function (ISMF) represented by a filled square: component having this class of management function always acts as managing entity. ASB

C

CSB CP1 MNO

CP2

ISP

management link physical link

-- Component Management Function (CMF) -- sub-Integral Service Management Function (sub-ISMF) -- Integral Service Management Function (ISMF)

Fig. 2. Service Management Topology

In addition, Table 1 provides brief descriptions of the management functions and management information exchanged between managing en managed entities. The table describes how components interact and what service session management information should be exchanged. In the chosen management model, ASB is assumed to have the information about which CP it should contact. ASB negotiates the requested content and the corresponding QoS level with both CP1 and CP2. There is a case where both CP1 and CP2 agree to provide the service component, then ASB will need to make a choice between one of the two service components offered. This situation is conceivable in practice when the prices of the service offered by CP1 and CP2 differ. The Service Component Provisioning is managed at local level, this means that if an actor confirms to the integral- or sub-integral management that the requested service component can be provided and service component creation confirmation message is received from the integral- or sub-integral management, then the actor must ensure a continuous provisioning of the service component (e.g. the final mapping of service confirmation message to network elements). These kinds of local management functionality are treated as ìinternalî management activities of a service component, which are not relevant for the modeling of service session management.

Formal Modeling of Service Session Management

Component

Management Function

Management Information Exchange

Interact with

- send service request

- content M + Q1 or Q2

- ASB

- terminate service

- termination message

- ASB

- negotiate the requested content and the corresponding QoS level

- content M + QoS level

- CP1 and CP2

- negotiate the required network resource a the end-to-end transmission of the content

- QoS, source CP1 and CP2

- CSB

- select service components and integrate these to obtain an end-to-end application service

non

non

- inform C if the requested service can be delivered or not

- service confirmation message

-C

- create connectivity and application service provisioning

- service creation message

- CSB, CP1 or CP2

- abort service provisioning when needed

- service return message

- CSB, CP1 or CP2

- terminate service

- service termination message

- CSB and CP1 or CP2

- respond to ASB if the requested content and the corresponding QoS is available

- content M + Q1 or Q2

- ASB

- negotiate network resource to destination C

- service negotiation message

- MNO

- negotiate network resource from source CP1 and CP2

- datin1, datin2, Q1 and Q2

- ISP

- respond if the requested endto-end connectivity can be provided or not

- service confirmation message

- ASB

- create connectivity service provisioning

- service creation message

- MNO and ISP

- service confirmation message

- CSB

MNO

- respond to CSB if the requested connectivity service can be provided or not

- service confirmation message, Q1 or Q2, datin1 or datin2

- CSB

ISP

- respond to CSB which connectivity service can be provided

C

ASB

CP1, CP2

CSB

Table 1. Management Functions and Management Information Exchange

41

42

M. van Le, B.J.F. van Beijnum, and B.L. de Goede

4 Modeling Method and Tools 4.1 Modeling Methodology State of the Art formal modeling methods can be applied to the previously presented service scenario and its management. More recently Architectural Description Languages have emerged. In review, the modeling concepts and tools relate as shown in Figure 3. TOO LS

CO NCEPT

Architecture Description Langua ge

St ru ctu ra l M od eli ng

F orm al Specification Language

Beh av ior al Mo de lin g

Mo del Checker

O ver al l M od el An aly sis

Fig. 3. The Modeling Concept

First, an Architecture Description Language (ADL) is used to express the structural aspect of a service architecture. The structural modeling step describes entities involved in an architecture and how they are connected to each other, through which interface, etc. A structural model should represent the topology of service relationships between the involved actors. Moreover, a structural model should contain the following information: • Which actor is responsible for which class of service component; • Through which interface should a certain service be provided; • How end-to-end service is provisioned; • Clear distinction of different class of interactions (e.g. basic signaling, accounting signaling, data stream, etc.); In the area of software engineering, ADLs have been identified as promising approach to bridge the gap between requirements and implementation in the design of complex systems [5]. A few well-known ADLs among system designers are: Darwin, Meta-H, Rapide, Wright. In addition, some of the ADLs support the capability to specify dynamic structures, a capability needed to capture roaming users. Second, the structural model is enriched with a behavior specification using a Formal Specification Language (FSL) that enables the behavioral modeling. The behavior of each component is specified in terms of process algebra. Third, to gain confidence in the correctness of the system, Model Checker is used once the behavioral model has been specified. Model checking techniques are especially aimed at the verification of reactive, embedded systems, i.e. systems that are in constant interaction with the environment. These techniques allow the designer to ver-

Formal Modeling of Service Session Management

43

ify that a mathematical model of a system satisfies logical properties depicting correct system behavior [6]. The followings summarize important aspects of model analysis: • Consistency: does the behavior of an entity conform the protocols, which are being used to interact to other entities? • Completeness: is any part missing in the system? (or, could we remove some parts and still being able to make sure that the system still function properly?) • Refinement: can one architecture be substituted for other? • Verification: does an implementation conform the architecture? • Performance and Reliability: can the server handle the expected demand? will average response time meet the requirements? which entity is the bottleneck? how would server/database replication affect the serverís performance? The model-checking step does not only help to analyze the behavioral model, it also provides good comprehension of the system in order to arrive at the management design. 4.2 Languages and Tools Most of ADLs have some formal semantics, which allows model checking. Darwin uses Finite State Processes (FSP), whereas Wright uses Communicating Sequential Processes (CSP). Both FSP and CSP are formal specification languages. Meta-H and Rapide use their own specification languages [7, 8, 9, 10]. Although limited in the number of states that can be captured and analyzed, Darwin/FSP is easy to use and therefore well suited as initial choice for this study. Darwin Architecture Description Language has been successfully applied in the area of software engineering to describe the structure of distributed telematics systems [11]. Darwin is a declarative configuration language, which may be used to describe the overall architecture of a distributed system as a hierarchical composition of interconnected components. Each Darwin component is an instance of a component type. The component interfaces are defined in terms of provided or required service. Components are connected by a fixed set of connector types. The effect of binding a provided service with a required service that corresponds to the composition of the subcomponents, and a configuration corresponds to the parallel composition of the components. Darwin has the ability to specify dynamically changing architectures: new components may be created as the result of the invocation of a particular kind of provided service [12, 13, 14]. To obtain the Darwin description of the Content Retrieval Service as illustrated in Figure 2, we make use of the Software Architect's Assistant (SAA) [15], a visual programming tool that supports the construction of distributed programs through the design of their architectures. SAA generates Darwin description from a graphical representation of the system. Given the overall structure specified in Darwin, this Darwin description can be parsed to an FSP specification using FSP's composition constructs. The behavior of each component can be specified in FSP. The combination of the two gives the entire system specification in FSP. For overall model analysis the tool LTSA is used [16].

44

M. van Le, B.J.F. van Beijnum, and B.L. de Goede

5 Modeling of Service Management 5.1 Modeling of the Structure of the Content Retrieval Service From the structural view, each involved actor is modeled as an independent component. Interaction between two components takes place through ìbindingî. Distinction between two types of bindings is made explicitly, namely: signaling binding and data binding. Further, we chose to model unidirectional bindings for clarity purposes of management information exchanging. Figure 4 depicts the graphical representation of the service architecture and the corresponding Darwin description. Management information exchange is enabled by unidirectional signaling bindings, ìsig1_nameî indicates an initiated interaction of a component, whereas ìsig1_nameî indicates an absorbed interaction of that same component.

sig1_ cp1 s ig1 _c

sig1_ c sig0_ cp1

sig0_c sig0_ cs sig1_ cs

sig0 _c p1

CP1

ASB

C sig0_ c

sig1_c p1

s ig1 _c p2

s ig1_ cp2

sig0_ cp2

sig0_ cp2

CP2

datin

datout2

sig0_ cs sig1_ cs datout

E2ECB

da tin2

da tin1

da tout1

compo nent SYSTE M { inst asb : ASB ; cp1 : CP 1; cp2 : CP 2; e2ecb : E2E CB; c : C; bind e2ecb. sig_c s0 -- asb. sig_c s0; asb.si g1_cs -- e 2ecb. sig1_ cs; e2ecb. datin 1 -- cp1.d atout ; e2ecb. datin 2 -- cp2.d atout ; asb.si g1_cp 1 -- cp.si g1_cp 1; asb.si g1_cp 2 -- cp.si g1_cp 2; c.dati n -- e2ecb .dato ut; cp1.si g1_cp 1 -- asb.s ig1_c p1; cp2.si g1_cp 2 -- asb.s ig2_c p1; asb.si g0_cp 1 -- cp1.s ig0_c p1; asb.si g0_cp 2 -- asb.s ig0_c p2; c.sig1 _c -- asb. sig1_ c; asb.si g0_c -- c. sig0_ c;}

Fig. 4. Compositional Service Architecture

The refinement of the end-to-end Connectivity Broker component (E2ECB) is shown in Figure 5. The choice to model CSB, MNO and ISP as sub-components has to do with the separation of concern in terms of provided service components. In doing so, ASB delegates a portion of management tasks to CSB that needs to take care of the end-to-end connectivity service. The CP1 and CP2 offer their application services physically through data interfaces datin1 and datin2, respectively. Once the data arrives at ISP, it is transmitted through one single data binding in the direction of MNO and then C. To summarize, the model contains the following parameters that govern the number of end-to-end service provisioning scenarios and therefore the number of states: • 1 content in a single segment: M • 2 content providers: CP1 and CP2 • 2 QoS levels: Q1 and Q2 • 2 data interfaces: datin1 and datin2

Formal Modeling of Service Session Management

s ig 0 _c s

s ig1 _ c s

E2ECB s ig1 _ c s

s ig0 _ c s

CSB

s ig0 _ m

s ig1 _ m

s ig0 _ is

s ig0 _ m

s ig 1 _m

s ig0 _ is

M NO

s ig1 _ is

s ig 1 _ is

IS P

da t out

d at in2 da tin

da to ut

da to ut

da tin2

d a tin1 da t in1

45

co mpone nt E2 ECB { p rovid e sig1 _cs; dato ut; r equir e sig0 _cs; dati n1; dati n2; i nst c sb : CSB; m no : MNO; i sp : ISP; b ind c sb.si g0_m -- mn o.sig0 _m; c sb.si g0_is -- i sp.sig 0_is; s ig1_c s -- csb.s ig1_cs ; c sb.si g0_cs -- s ig_cs0 ; i sp.da tin1 -- da tin1; i sp.da tin2 -- da tin2; d atout -- m no.da tout; m no.si g1_m -- cs b.sig1 _m; i sp.si g1_is -- c sb.sig 1_is; m no.da tin - - isp .datou t;}

Fig. 5. Refinement of E2ECB component

5.2 Specifying the Behavior of the Service Components The system behavior is modeled as a Labeled Transition System (LTS) with shared actions to synchronize processes. It is important to keep in mind the recurrence of the processes running in a system during the behavioral modeling phase, where ìlifeî is created within each primitive1 component. In terms of finite states processes, once a service session life cycle is terminated, all processes must return to the initial state. The behavioral model of CP1 is shown in Figure 6 together with its corresponding FSP specification. The result of the systemís overall behavior is a composite LTS: || SYSTEM = (asb:ASB || cp1:CP1 || cp2:CP2|| e2ecb:E2ECB || c:C) where

|| E2ECB = (csb:CSB || mno:MNO || isp:ISP)

The behavior of primitive component CP1 is relatively simple compared to other components in the model. CP1 is illustrated here as example to provide an idea about how the behavior of a primitive component can be described and graphically displayed as a finite state machine using LTSA. LSTA provides two principle options to analyze a system: safety- and progress analysis: • Safety Analysis checks systemís deadlock, which is generally a bad state from which no further actions can be executed. Deadlock might occur in case a component cannot finish it session life cycle before the start of another session. • Progress Analysis checks systemís property to ensure that whatever state the system is in, there is always the case that a specified action will eventually be executed. In addition to basic analysis options, one can also check system ës property with some pre-defined requirements. For instance, the responsiveness property of an accounting system ensures that accounting must stop whenever provision failure occurs. 1

without decompositions capability

46

M. van Le, B.J.F. van Beijnum, and B.L. de Goede asb.sig0_cp1.negCont[21][11]

asb.sig0_cp1.negCont cp1.sig [21][12] 1_cp1.conte asb.sig nt[21] 0_c [1] p1.provision[2 cp1 1][12 .datout ] 1.send[21][1 asb.sig0_c 2] p1.playing[21]

cp1:CP1

0

1

2

3

4

cp1.sig1_cp1.cont asb.si ent[2 g0_ 1][1] cp1.provision cp1.da [21][11] tout1.send[21 asb.sig0 ][11] _cp1.playing[21]

5

6

7

8

9

10

cp1.sig1_cp1.content[21][0] asb.sig0_cp1.return

cp1.sig1_cp1.enddat[21] cp1.sig1_cp1.content[21][0] asb.sig0_cp1.return

cp1.sig1_cp1.enddat[21]

// Conte nt Pr ovider 1 CP1 = (s ig0_c p1.neg Cont[ m:Mov ies][ q:QoS ] -> (s ig1_c p1.con tent[ m][Fa lse] -> CP 1 |s ig1_c p1.con tent[ m][Tr ue] - > (s ig0_c p1.ret urn - > CP1 |s ig0_c p1.pro visio n[m][ q] -> dato ut1.s end[m] [q] - > si g0_cp 1.play ing[m ] -> sig1_ cp1.e nddat [m] -> CP1) )).

Fig. 6. Content Providerís behavioral model

In order to gain confidence in the correctness of the model, we first run the reachability analysis to check for deadlock in the system. A summary of the reachability analysis is given in Figure 7. Compiled: ASB; Compiled: CP1; Compiled: CP2; Compiled: CSB; Compiled: MNO ;Compiled: ISP Composition: E2ECB = csb:CSB || mno:MNO || isp:ISP State Space: 107 * 11 * 39 = 2 ** 17 Composing...; -- States: 221 Transitions: 342 ... Compiled: E2ECB Compiled: C Composition: SYSTEM = asb:ASB || cp1:CP1 || cp2:CP2 || e2ecb:E2ECB || c:C State Space: 216 * 11 * 11 * 221 * 11 = 2 ** 28 Analysing... Depth 32 -- States: 1063 Transitions: 2036;Memory used: 3648K No deadlocks/errors

Fig. 7. Reachability Analysis

As indicated, there is no deadlock in the system. However, reachability does not guarantee that a systemís behavior completely meets the systemís requirements. Therefore, all possible service scenarios must be examined with respect to expected management functionalities of each component. According to the systemís degrees of freedom, ASB must manage in total 64 operational service scenarios, namely: [M:1]x[QoS:2]x[CP:2]x[Availability:2]x {[DataInterface:2]x[QoS:2]x[Availability:2]} = 64 Where [Availability] is a YES or NO response of a managed entity in responding to a managing entity about the availability of the requested service component.

Formal Modeling of Service Session Management

47

5.3 Example of End-to-End Service Provisioning As mentioned previously, the behavioral modeling results in 64 operational service scenarios. Figure 8 provides the message sequence diagram of a scenario to illustrate how the Service Session Life Cycle mentioned in section 2 can be followed to obtain an end-to-end service provisioning. user Service Session Invocation

ASB

CP1 2 7

Service Session Negotiation

CSB

ISP

MNO

3

4

8

5

6

9

10

11 12

Service Session Creation

13

14

15

17

Service Session Provision Service Session Termination

CP2

1

16 18 19

20

21

22

23

24

Fig. 8. Message Sequence Diagram of a service-provisioning scenario

• Step 1 ñ The user invokes a service. • Step 2 and 3 ñ ASB negotiates the requested content and the corresponding QoS. • Step 4 ñ ASB negotiates the required network resource for the end-to-end transmission of the content. • Step 5 and 6 ñ CSB negotiates network resource at ISP and MNO • Step 7 and 8 ñ CP1 and CP2 respond positively to ASB. • Step 9 and 10 ñ ISP and MNO respond positively to CSB. • Step 11 ñ CSB responds positively to ASB about the availability of the end-to-end connectivity service. • Step 12 ñ ASB creates the application service component offered by CP1. • Step 13 ñ ASB informs CP2 that it the offered service component is dropped. • Step 14 ñ ASB creates the end-to-end connectivity service offered by CSB. • Step 15 and 16 ñ CSB creates the connectivity service components offered by ISP and MNO. • Step 17 to 19 ñ The actual end-to-end service is being provisioned. • Step 20 ñ The user terminates the ongoing service. • Step 21 and 22 ñ ASB terminates the application and end-to-end service components. • Step 23 and 24 ñ CSB terminates the connectivity service components.

48

M. van Le, B.J.F. van Beijnum, and B.L. de Goede

6 Conclusions and Future Work In our study we focus on the service management aspects, we abstract from the physical relations by looking at the service relations between the actors. We decompose telematics service management in a multi-provider environment in such a way that application of formal method results in an operational relevant behavioral model. The proposed concept serves as a basis for further application of formal method in the area of Service Management to investigate crucial properties of end-to-end Service Management Systems in order to arrive at a secure and optimal management system design. As far as the future work is concerned, we aim at exploring the possibility to include mobility aspects in the modeling of multi-domain Service Management. As mentioned, Darwin has the ability to specify dynamically changing structures, which can be useful to describe the mobility of the customer when roaming from one to another domain and the systemís behavior as a result of the userís mobility. Future extensions of the work presented here include also the modeling of multi-domain Accounting Management and property verification of service session management and its associated session accounting management.

Reference 1. Shingo Ohmori , ìThe concept of future generation mobile communication based on broadband access technologyî, IEEE Communication Magazine, Dec. 2000. 2. B.L. de Goede, ìOperational Management of Telematics Systems and Servicesî, ISBN: 90365-1709-5, Dec. 2001. 3. Lewis, D.A.: "A framework for the development of service management systems for the open service market", PhD thesis, UCL, Mar. 2000. 4. Q. Kong et al: ìA Management Framework for Internet Servicesî, CiTR technical journalVol.3, 1998. 5. J. Kramer, J. Magee, ìExplosing the skeleton in the Coordination Closetî, Proc. of the 2nd International Conference on Coordination Models and Languages, Sept. 1997. 6. R. Allen: ìA Formal Approach to Software Architectureî, PhD thesis, 1997. 7. ìRapide Overviewî, http://pavg.stanford.edu/rapide/overview.html 8. îThe Wright Architecture Description Languageî, http://www-2.cs.cmu.edu/afs/cs/ project/able/, Canergie Mellon - School of Computer Science,1998. 9. ìDarwin - An Architecture Description Languageî, http://www.doc.ic.ac.uk/~jsc/research/, Dept. of Computing, Imperial College, Apr. 1997. 10. ìMeta-H Language and Toolsî, http://www.htc.honeywell.com/ projects/dssa/dssa_tools. 11. H. Fossa: ìImplementing Interactive Configuration Management for Distributed Systemsî, International Conference on Configurable Distributed Systems, Annapolis MD, May 1996. 12. J.Kramer: ìAnalysing Dynamic Change in Software Architectures - A Case Studyî; 4th International Conference on Configurable Distributed Systems, Annapolis, May 1998. 13. N. Dulay, ìDarwin Refererence Manualî, Dept. of Computing, Imperial Collegeî, 1994. 14. J. Magee, N. Dulay, J. Kramer, ìSpecifying Distributed Sofware Architectureî, Proc. of the 5th European Software Engineering (ESECí95), Sept. 1995. 15. K. Ng, et al: ìA Visual Approach To Distributed Programmingî, in Tools and Environments for Parallel and Distributed Systems , A. Zaky and T. Lewis (Eds.), Kluwer Academic Publishers, Feb. 1996. 16. J. Magee, J. Kramer: ìConcurrency: State Models & Java Programsî, Wiley, 1999.

Network Requirement for Management of Multimedia over Wireless Channel Bing Zheng1 and Mohammed Atiquzzaman2 1

New Focus Inc 5215 Hellyer Ave., San Jose, CA 95138. Email:[email protected] 2 School of Computer Science University of Oklahoma, Norman, OK 73019-6151. Email: [email protected] Abstract. Wireless channels have a high channel bit error rate and limited bandwidth. The high bit error rate degrades the quality of multimedia transmission, and also increases the buffer size requirement at the client. In this paper, we propose a novel selective retransmission scheme for multimedia transmission over wireless ATM networks using the ABR service. Our objective is to develop a simple cost effective scheme which offers quality of service for MPEG compressed video transmitted over a noisy wireless channel. We have developed an analytical model to determine the networking requirements and receiver buffer size for our proposed scheme for transmitting multimedia over wireless ATM.

1

Introduction

The transmission of multimedia over wireless channel is becoming a research topic of growing interest [1,2,3]. With the emergence of next generation high speed wireless ATM networks, it is expected that multimedia will be the main source of traffic in future wireless ATM networks. Wireless transmissions use radio as the transmission media which is easily affected by environments such as buildings, moving objects, and atmosphere, etc., giving rise to issues such as effective bandwidth allocation. The ATM Forum has standardized four service classes: Constant Bit Rate (CBR), Variable Bit Rate (VBR) including real time VBR and non-real time VBR, Available Bit Rate (ABR) and Unspecified Bit Rate (UBR). CBR has the highest quality of service guarantee and the highest service cost. However, ABR offers cost effective service at an acceptable quality of service, flexible bandwidth allocation, and has a high utilization of network bandwidth. We will therefore, investigate the effectiveness of using the ATM ABR service to transport multimedia over a wireless ATM network. By combining the advantages of ATM and wireless, it is possible to implement a wireless ATM network which will offer high speed, quality of service, and mobility to future multimedia users [4,5,6]. Networked multimedia applications include remote education, video-ondemand, tele-shopping, home game entertainment, etc. Because of traffic characteristics such as high bit rate, video will be the dominant traffic in multimedia The original version of this chapter was revised: The copyright line was incorrect.This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 49–61, 2002. c IFIP International Federation for Information Processing 2002

50

B. Zheng and M. Atiquzzaman

streams, and hence needs to be managed efficiently [7]. Video must be compressed (for example, MPEG) to reduce the required transmission rate. In contrast to data, multimedia can tolerate a certain level of error. Therefore, although a wireless network has a high bit error rate when compared to a wireline network, it is possible to cost effectively transmit multimedia over wireless networks with acceptable quality of service (QoS). In their experimental studies on effective algorithms for video transmission over a wireless channel [8], the authors studied the go-back-W retransmission scheme with window size W. They also considered the receiver buffer size for single retransmission. Their experimental results have shown that retransmission without FEC can improve the video transmission quality. However, their studies did not present detailed theoretical model and analysis of their retransmission scheme. Since the effectiveness of a retransmission scheme depends on the network parameters such as channel bit error rate, channel transmission rate, and video parameters such as the video frame size, it is very important to study the relationship between these parameters and the effectiveness and efficiency of retransmission schemes. Moreover, to sustain continuous video display at the receiver considering transmission over a noisy channel, continuous monitoring of the network status is required to prevent transmission collapse. Finally, during the time taken for retransmission, the receiver can only decode the buffered data upto the errored frame. To prevent starvation at the receiver (resulting in frozen display), the receiver buffer must have a minimum fill level to sustain continuous display during retransmission. The minimum fill level thereafter decides the minimum receiver buffer size. The longer the retransmission time, the larger the minimum fill level. Since the retransmission time depends on the channel bit error rate, the minimum receiver buffer size depends on the channel bit error rate. The authors in [9] have reviewed several error resilient video coding methods for MPEG-2 transmission over wireless ATM networks. They found that in a WAN testbed, the cell losses and wireless packet losses occur independently. The high bit error rate of wireless channels raises a number of challenging issues for multimedia transmission over a wireless ATM network. Such issues include techniques to ensure a continuous display at the receiver in spite of errors in the received video. Currently, errors are compensated mainly by Forward Error Correction (FEC) which employs redundant bits to passively correct errors at the receiver. Although FEC can correct some errors without retransmission, it needs extra bandwidth to transmit a lot of redundant bits for correction. This lowers the utilization of bandwidth which is expensive for wireless channels. Furthermore, implementing FEC requires a lot of computing power at the receiver, thereby increasing the cost of the receiver. Additionally, complicated error concealment algorithms will have to be implemented at the receiver for different error types [10]. All of the above issues may prevent cost effective wide scale commercial deployment of wireless multimedia receiver if FEC is used at the receiver. Therefore, new schemes for error correction at the receiver are required for widespread deployment of cost effective multimedia transmission over noisy wireless links.

Network Requirement for Management of Multimedia over Wireless Channel

51

In this paper, a new selective retransmission scheme is proposed for multimedia transmission over noisy wireless channels in order to ensure acceptable video quality at the receiver and allow the design of a cost effective receiver. Since multimedia can tolerate a certain level of error with little effect on picture quality, our scheme sets a threshold on the number of errored cells, and to decide whether retransmission is required. To prevent the receiver from starvation (i.e. ensuring a continuous display), a receiver buffer is required. Our objective is to develop an efficient error handling scheme for wireless transmission of multimedia, and to show that the scheme results in a cost effective wireless receiver. The significance of our scheme is that it does not require powerful computing capability and complicated error correction algorithm to be implemented at the wireless receiver. The wireless receiver will require simple hardware and thereby reduce the receiver cost. Our scheme also offers a method to monitor the status of the transmission system to check if the conditions to maintain transmission are satisfied. Our performance measure is maintaining a continuous display at the wireless receiver. The rest of the paper is organized as follows. In Section 2, we propose the system stack model and our proposed selective retransmission scheme. Analysis of a multimedia wireless system using our proposed scheme is performed in Section 3. In Section 4, the minimum receiver buffer size is derived. Numerical results obtained from our analysis are given in Section 5, followed by conclusions in Section 6.

2

System Model and Proposed Retransmission Scheme

In this section, we describe the system model and the operating principle for our selective retransmission scheme. It will be used in Section 3 to derive values of network parameters and the size of receiver buffer to sustain continuous display at the receiver. As shown in Figure 1, the system consists of a multimedia server and receivers connected by a wireless ATM network using the ATM ABR service. The protocol stack consists of four layers as shown in Figure 2. MPEG compressed video consists of I, P and B frames. The frames are grouped to form a special structure called Group of Picture (GoP). Each GoP includes an I frame followed by a number of P and B frames. A GoP is denoted by M mN n which represents a total of n frames in the GoP with m − 1 number of B frames between the I/P or P/P frames. At the decoder, error in a received I frame will affect all the frames in the GoP, while error in a received P frame will affect the quality of the subsequent P and B frames. An error in a received B frame only affects its own quality. Therefore, I frame is the most important, while the B frame is the least important. To maintain the quality of video at the receiver, we propose the following retransmission scheme based on errors in the received video stream at the receiver. – If the number of errored cells in an I frame is higher than an acceptable error threshold for the I frame, the cells belonging to the I frame are discarded, and the I frame is retransmitted from the multimedia server;

52

B. Zheng and M. Atiquzzaman

ATM Central Station ATM

ATM Link

ATM Central Station

ATM Link

Base Station Wireless multimedia

ATM Link

ATM Link

receiver

ATM Base Station Wireless multimedia

Multimedia Server

MPEG

MPEG

AAL

AAL

ATM

ATM

PHY

PHY

receiver

Fig. 1. The element of wireless ATM network for multimedia application.

Fig. 2. The protocol stack for MPEG video transmission over ATM wireless networks.

– If the number of errored cells in a P frame is higher than the acceptable error threshold for the P frame, the P frame is discarded and is retransmitted; – Errored cells in B frames are discarded and are not retransmitted. In the next section, we develop a mathematical model to analyze our proposed scheme, to derive the network requirements to sustain a continuous video display at the receivers, and to determine the minimum size of the receiver buffer. In our discussion, we assume: – The wireless channel has a bit error rate which is independent of the traffic pattern and the channel transmission rate; – Each bit has equal probability of being errored during transmission. Moreover, each cell has equal probability of having errors; – The channel has fixed round trip time between the server and receiver; – The ATM ABR service is used to transmit multimedia [7], i.e. the amount of bandwidth available to the application changes over time depending on the network congestion level.

3

System Analysis

In this section, we develop a mathematical model to analyze our proposed retransmission scheme. First, we define the following variables. 3.1

Notation

ρ: Channel bit error rate of the wireless link; SI , SP and SB : Average size of I, P and B frames respectively of MPEG video; N : Number of bits in the payload of a cell (For an ATM cell, N =48*8); NI , NP and NB : Average number of cells belonging to I, P and B frame in a GoP;

Network Requirement for Management of Multimedia over Wireless Channel

53

dI , dP : Threshold for the acceptable number of errored cells in I and P frames respectively; p: Probability that a cell has error; WI : Probability that an I frame is errored given that an error has occurred; WP : Probability that a P frame is errored given that an error has occurred; PI (dI ): Probability that an I frame needs to be retransmitted, called I frame transmission failure probability; PP (dP ): Probability that a P frame needs to be retransmitted, called P frame transmission failure probability; Td : Fixed round trip time (FRTT) from server to receiver; Tr : Average successful retransmission time for an errored frame, called the average system recover time interval. This is the average time required for successful retransmission; τf : Average time interval between consecutive frame errors, called average system failure time interval; λ: Channel transmission rate; µ(t): Multimedia playback rate at time t; Cmin : Minimum receiver buffer size. 3.2

Average System Failure and Recover Time

Since the probability of a bit being in error is ρ, the probability that the payload of a cell has no error is (1 − ρ)N . Therefore, the probability p that a cell is in error can be expressed as: (1) p = 1 − (1 − ρ)N The probability PI (dI ) that an I frame needs to be retransmitted is the probability that the number of errored cells in the frame is higher than the threshold dI . Therefore, PI (dI ) can be expressed as: PI (dI ) = 1 −

dI  

NI i

 i p (1 − p)NI −i

(2)

i=0

dI NI  i NI −i represents the probability of the I frame having where i=0 i p (1 − p) upto dI errored cells. Similarly, PP (dP ) can be expressed as: PP (dP ) = 1 −

dP  

NP j

 j p (1 − p)NP −j

(3)

j=0

Given that a frame is in error, the probability that the frame is an I frame is proportional to the size of the I frame in a GoP. Therefore, WI can be expressed as: SI WI = (4) n SI + SP ( m − 1) + SB m−1 m n

54

B. Zheng and M. Atiquzzaman

where the denominator represents the size of a GoP. Similarly, we can obtain the expression for WP as: WP =

SI +

n SP ( m

SP − 1) + SB m−1 m n

(5)

If an I frame is in error, the average time TrI for successful retransmission is expressed as: TrI = (Td + τsI )(1 − PI (dI )) + 2(Td + τsI )PI (dI )(1 − PI (dI )) + ...... Td + τsI = 1 − PI (dI )

(6)

where τsI = SλI is the time required to transmit an I frame into the wireless channel. Similarly, TrP for successful retransmission of a P frame can be expressed as: Td + τsP (7) TrP = 1 − PP (dP ) where τsP = SλP is the time to transmit a P frame into the channel. Therefore, the average system recover time (Tr ) is the statistical sum of the I and P frame recover times TrI and TrP with weights WI and WP respectively. Therefore, Tr can be expressed as: Tr = TrI WI + TrP WP

(8)

Because a wireless channel has significant error rate, let’s say that an error happens after time τf which is defined as the average system failure time. Since the acceptable error thresholds are dI and dP for I and P frames respectively, the average system failure time is: τf =

min(dI , dP ) λρ

(9)

where the λ is the available bandwidth of the ABR connection. An interactive receiver can perform VCR like interactive functions such as stop, playback, fastforward (FFW), and fastbackward (FBW) which are represented by a state transition diagram as shown in Figure 3. In the FFW/FBW state, the receiver will consume data at a higher speed than the playback state. Therefore, in response to a FFW/FBW request, the sever will request higher bandwidth than is required in normal playback. The bandwidth λ required by the server is decided by the level of interactivity of the receiver and the available bandwidth. Let’s denote the stop, playback, FFW and FBW states of the receiver by 0, 1, 2, 3, and the state transition probability from state i to state j (where 0 ≤ i, j ≤ 3) is represented by θij . If stationary state probability vector for the

Network Requirement for Management of Multimedia over Wireless Channel

55

receiver is X = (X0 , X1 , X2 , X3 ), X can be obtained by solving the following equation [11]: 1 1 + θ0,1 /θ1,0 (1 + θ1,2 /θ2,1 + θ1,3 /θ3,1 ) θ0,1 X1 = θ1,0 (1 + θ0,1 /θ1,0 (1 + θ1,2 /θ2,1 + θ1,3 /θ3,1 )) θ0,1 θ1,2 X2 = θ1,0 θ2,1 (1 + θ0,1 /θ1,0 (1 + θ1,2 /θ2,1 + θ1,3 /θ3,1 )) θ0,1 θ1,3 X3 = θ1,0 θ3,1 (1 + θ0,1 /θ1,0 (1 + θ1,2 /θ2,1 + θ1,3 /θ3,1 )) X0 =

(10)

For interactive multimedia, the expected channel transmission rate E[λ(t)] is therefore expressed as the statistical sum of the playback rate, fastforward rate, and fastbackward rate with weights X1 , X2 , and X3 . E[λ(t)] = [X1 + k(X2 + X3 )]µ(t)

(11)

where k is the FFW/FBW speed factor implying that the channel transmission rate is k times that of the normal playback rate. 3.3

System Model and Dynamics

If the receiver receives a frame with the number of errored cells exceeding the threshold, the server will retransmit the corresponding frame by switching to the retransmission state. Therefore, the status of the system can be divided into two states as shown in Figure 4: – State 0: the normal state in which multimedia is continuously received at the receiver; – State 1: the retransmission state in which the system is retransmitting a frame. The receiver consumes data upto the errored frame and then waits until the successful arrival of the retransmitted frame.

1

Playback

τf Fastforward

Fastbackward

Stop

Fig. 3. The interactive operation state at receiver.

1-

1

τf

1

0

1 1- T r

1 Tr

Fig. 4. The state diagram of the serverclient type wireless multimedia transmission system.

56

B. Zheng and M. Atiquzzaman

From Figure 4, the state transition matrix M is expressed as:   1 1 − τ1f τf M= 1 1 − T1r Tr

(12)

Let V = (V0 , V1 ) denote the long term stationary state vector, which satisfies the stationary equation: V = VM (13) By solving the stationary equation, the steady state probabilities are obtained as: 1 1+α α V1 = 1+α V0 =

(14) (15)

where α = Tτfr . In order for the system to work satisfactorily, it should be at state 0 most of the time, i.e., α xíi) then yíi =xíi-1; if (yíi = xíi=0) then xíi =xi; yíi =yi; For any packet misses its deadline: If(xíi>0) then xíi =xíi-1; yíi =yíi -1; if(xíi = yíi=0) then xíi = xi; yíi = yi; else if(xíi = 0) then if(xi > 0) then yíi = yíi + (yi -xi)/ xi ; if(xi = 0) then yíi = yíi + yi;

Figure 1: Pseudo-code for packet processing (from [1])

R.West, Chris. Poellabauer [3] fully implemented DWCS on Linux machines. This implementation uses a circular queue for each stream. Packets of the same stream are added to the tail of the queue and removed from the head of the queue. Two heaps are set up for deadline and loss-tolerance. With the heaps, packet insertion can be done using heap sorting. Packet precedence comparison will follow the rules of Table 1. DWCS checks streams for packets that have missed their deadlines every time a packet has been serviced. In the worst case, every stream can have late packets when the scheduler completes the service of one packet. The deletion of these packets from the losstolerance heap and the heap adjustment will cost O(nlogn). Also, O(logn) time is required to find the next packet to be sent from the n streams. Thus, the current implementation on Linux with heap structure is not scalable as the number of active streams in the system increases. When the stream rate is low (more active streams coexist in the system) or the bandwidth is high compared to the stream rate, the two heaps will be adjusted frequently. Since more time is spent on insertion and selection, there will be further degradation in the throughput as the system is overloaded.

4 4.1

IXP Architecture IXP Hardware

The IXP (Internet Exchange Processor) 1200 in Figure 2 is a network processor from Intel. The processor consists of a StrongArm core, which functions as a traditional microprocessor. Connected to this core are six microengines (RISCs), which are responsible for managing the network data. Both the Strong Arm core and the six microengines run at the same clock rate (232 MHz). Furthermore, all microengines connect to a high-speed bus (IX bus, 80 MHz) for receiving and transmitting packet data. On the other side of the IX bus are external MAC devices (media access controllers). The MAC devices can be either multi-port 10/100BT Ethernet MAC, Gigabit Ethernet MAC, or ATM MAC. Both the StrongArm core and the microengines can access all of the address space of the SRAM and the SDRAM.

188

4.2

W. Shi et al.

Pipelined Processing

There are two choices in assigning microengines (µEs) to tasks. One is to duplicate the functionality of each µE, so that all µEs work in the same way to service the ports in a round robin fashion. The second choice is to assign the microengines with different tasks, so that each of them can work on one job. The second choice allows threads in different microengines to be chained to form a processing pipeline.

Figure 2: Block Diagram of IXP network processor

5 5.1

Hierarchically Indexed Linear Queue (HILQ) ñ Data Structure and System Design Drawbacks of the Heap Structure

The current Linux implementation of DWCS has many drawbacks that make it unsuitable for network processors: (i) The heap needs many memory operations. Selection and insertion each require O(log(n)) time. In the worst case, when every stream may have packets that missed deadlines, O(nlog(n)) time will be needed to delete and rearrange the heap. (ii) The loss-tolerance heap will potentially become a bottleneck when packet insertion, packet dropping and packet selection happen concurrently. Introducing locks on the losstolerance heap will further degrade the performance. (iii) A centralized scheduler will surely become a bottleneck even with a small number of streams (5-10). To maintain Gigabit speed, a centralized scheduler should have a scheduling period within 500 ñ 600 microengine cycles (microengine runs at about 232MHz). 5.2

Assumptions

Several assumptions influence the HILQ data structureís design and implementation. They are as follows: (i) The IXP network processor can have very large off-chip memories, and a wellestablished memory hierarchy from scratchpad to SRAM, to SDRAM and it offers specific hardware for hash operations, freelists, memory locks, etc. So, a well-established data structure may benefit from this architecture.

Efficient Implementation of Packet Scheduling Algorithm

189

(ii) We choose a video/audio unified traffic, where the intervals between frames are in the range of tens to at most one hundred milli-seconds. In other words, the packet rate is as low as 20 ñ 60 packets/second for each stream. However, the number of streams can be extremely large (hundreds to thousands of streams). This paper addresses scalability in number of streams, but current implementation avoids the need of frame fragmentation and reassembly by scheduling MAC-layer packets. (iii) Our implementation deals with systems that are not overloaded, where the average stay time or the average delay time for each packet (including queuing time and processing time) is much less than the packet inter-arrival time, i.e. when the input and output packet rates are balanced. 5.3

Data Structures

Previous researches on priority queues [13][14][16][17] used heap or calendar queue data structures. Calendar queues have a good average performance, especially when the queue is sparse. Our design can be thought of as calendar queue-like, since the ìhash functionî is not arbitrarily defined but based on the deadline and loss-tolerance. To search for the next packet to be sent, we traverse the queue linearly. Figure 3 shows the overall data structureóthe Hierarchically Indexed Linear Queue (HILQ). Level 0 is a continuous queue divided equally into 100 segments. Each segment represents one milli-second, i.e. all the packets with deadline in this milli-second are stored into the segment accordingly. Currently, 100 milli-seconds (segments) are included, which means we only look forward for this amount of time. 100 milli-second is sufficient for stream data, where inter-arrival times are typically much smaller. Also notice that the segments are organized as a circular queue, in a sense that it will loop back after segment 99. The transmission microengine will work on the current milli-second segment and then go forward to the next milli-second segment when all packets in the current segment have been serviced. The receiving engine will put the packets directly into the queue according to their deadlines. For packets with deadline 100 milli-seconds in the future, this structure will cause problems. For example, if the current time is 100.050s, the transmission engine will work on segment 50. For a packet with a deadline of 101.020s, it will be put into segment 20. The transmission engine proceeds linearly till segment 99 and then loops back to segment 0. It will reach segment 20 at time 101.20s. If we insert a packet with a deadline of 101.60s, then it will be erroneously put into segment 60, i.e. it will be processed at time 100.60s. Thus, our implementation does not support arbitrarily disordered streams.

Level 0 segm ent 0

segm ent 1

segm ent 99

Transm ission Pointer

Figure 3: Hierarchically Indexed Linear Queue

190

W. Shi et al.

Within each segment, a number of slots are created which is actually a mapping from the values of x and y. Figure 4 shows the structure inside each segment. First, to avoid the calculation of x/y, an x/y table is used to retrieve the value of x/y via a single memory operation. The size of the table depends on the ranges of permitted x and y values. The current implementation uses 50 distinct values in the x/y table. These 50 values are from 0 to 1 with a step of 0.02. Each x/y value is rounded to the nearest one among these 50 predetermined values. In other words, we may have up to 0.01 inaccuracy during the round off. We consider this amount of inaccuracy tolerable in practice. In our experiments, x can have an integer value between 0 and 19; y can have an integer value from 0 to 29. Figure 4 shows the 1000 slots and the interpretation of each slot. Slots are ordered based on Table 1. The slots with smaller indices will be serviced earlier since they have higher precedence. If x/y values are identical, slots are ordered again by the value of x, the numerator. So, 20 slots for each of the 50 x/y values are added to the queue. Therefore, a total of 1000 slots are needed for each milli-second. Till this point, we have considered the precedence of x/y and x. However, it is still possible that packets with the same values of x/y and x fall into the same milli-second interval. In reality, this happens when streams are quite regular. We also observe that streams tend to cluster around several x/y values as x/y values are adjusted during the transmission of streams. This requests the existence of an element queue (a FIFO queue) for each slot. Actually, each slot currently contains information (header pointer and element number) about the element queue if there are packets in this slot. Multiple packets belonging to the same slot will be ordered by their arrival times and placed into appropriate queue position. (0.00,0)

(0.00,1)

(0.00,19

(0.02,0)

(0.02,1)

(0.02,19

Extended element queue 1

m elements

Y

1

2

1

1

2

0.5

3

0.33

3

(0.98,1)

(0.98,19

Ö

x/y table

k*m elements Basic element queue

(0.98,0)

x

X/Y

Figure 4: Elements inside a Segment

All three precedences have been taken into account in HILQ. To avoid frequent memory reallocation, there are two kinds of element queues for each slot, a basic element queue and an extended element queue. The basic element queue is created for a fixed number of packets, while the extended element queue is used only when the basic element queue is not large enough. Statistically, the extended element queue is rarely used. The two element queue structures can reduce memory needs while still providing enough support in the worst case. Memory usage is partly static and partly dynamic. The segments for each milli-second are allocated statically, whereas element queues (basic and extended) are dynamic data structures, which can be allocated on demand and released by the garbage collector.

Efficient Implementation of Packet Scheduling Algorithm

5.4

191

Pipeline Stages

The microengines are assigned three kinds of functionalityóreceiving, scheduling and garbage collection. They all operate on the shared HILQ data structure, while the element queues reside in the IXP SRAM for low memory access latency. We develop microcode component for each functionality and assign them to different threads to form a processing pipeline. 5.4.1 Receiving Threads When the network processor receives a packet, admission control will first check whether to accept it. Dropping packets by admission control is a violation of the DWCS rules. It can only happen when the system is running out of memory, which may occur when the system is overloaded. Receiving threads also do some common processing on the packet header like modification to the protocol fields, etc. After that, the pointer to the packet is inserted into a proper place in the queue. In DWCS, three property values are used to direct the packet insertionóD(deadline), x and y. Suppose the current time is Tnow(in milli-seconds), then the packet should be put into the segment with number (Tnow+D)%100. Then we will look up the value of x/y in the x/y table. In fact, the table stores the index number to the position inside the segment. Let the function LOOKUP do the mapping of the x/y table. We will insert the packet pointer into the element queue at the place LOOKUP(x/y)+x inside the segment. If the packet is the first one in a slot, a new basic element queue is allocated from the freelist. If the basic element queue is full at the time a new packet is inserted, an extended element queue is allocated from the freelist. 5.4.2

Scheduling Thread L e v e l 2 v e c to r in S R A M L in e a r q u e u e

2 n= N

0 1 2 3 . . . . 31 0 1 2 3 . . . . 31

L e v e l 1 v e c to r in sc ra tc h p a d

0 1 2 3 . . . . 31

0 1 2 3 . . . . 31

Figure 5: The Hierarchical Index Queue

The scheduling thread first works on the current milli-second segment and tries to send out all the waiting packets in this segment. The scheduling thread traverses the segment linearly. Certainly, there will not be any new packets put into the segment when the

192

W. Shi et al.

scheduling thread is working on it. Therefore, the search is unidirectional. We can also make use of the hierarchical index to skip empty positions quickly. When a packet is encountered, the scheduling thread will remove it from the queue, which includes resetting the pointer and releasing the element queue if needed. The x, y values of that stream should also be adjusted. After servicing the packets in the current milli-second segment, according to the DWCS algorithm, we should go forward to send the packets in the future segments (we should not wait in the current segment, otherwise the computation power is wasted). However, the future segments are dynamically changing, when new packets are inserted into their deadline slots. We cannot simply search forward. If a new packet belongs to an earlier place than the current search position, the scheduling thread has to go back and send it. Note that this will not degrade performance because when time for the next segment comes, the scheduling thread will move to the next segment immediately. All the packets before the next segment will be dropped. Figure 5 shows a two level hierarchical indexed queue that can further reduce search time (n=10, N=2n=1024). The level-one vector is a longword (4 bytes) in the scratchpad (on-chip) memory. The level two vectors are 32 longwords in the SRAM. Each bit of the level-two vector corresponds to an element in the queue. Each bit in the level-one vector corresponds to a longword in the level-two vector. To search out the next element with a packet, it requires only two memory reads. 5.4.3 Garbage Collection Thread The garbage collector is called every time the scheduler has moved to the next millisecond segment. It searches the processed milli-second segment, drops all packets in the segment, and adjusts the x/y values for the streams with dropped packets. Separate threads are used for garbage collection. Since scheduling and receiving threads will not work on the same data location as the garbage collector, we are sacrificing the accuracy of the DWCS algorithm. This is because, at the time garbage collector begins to work but packet-dropping has not been completed, the receiving threads are still using the old x/y values of the streams (those x, y values may be modified if packets are dropped by the garbage collector). The garbage collector can work extremely fast under light overload (a small number of packets are dropped), thus limiting the inaccuracy to a small range. When the system is under heavy load, packets will not be admitted by the receiving threads, automatically constraining potential inaccuracies.

6

Performance Evaluation and Analysis

We have implemented the HILQ algorithm with microcode. All experiments are conducted on the Intel-provided simulation environmentóIXP1200 Developer Benchmark 2.0. The IXP1200 workbench supports two microcode testing modes, simulation and real hardware. The user's microcode program can be tested in either of these two modes. It is capable of simulating most IXP devices and microengines with high fidelity. It also has a cycle accurate IX bus model that can simulate receiving and sending traffic data. We want to demonstrate that our HILQ based DWCS performs well with high-speed connections like Gigabit Ethernet. All the experiments are simulated in the workbench with a SRAM configuration of 8MB and SDRAM configuration of 64MB.

Efficient Implementation of Packet Scheduling Algorithm

193

To get an idea of how costly the SRAM operations are, we collected the data in Table 2 to show the distribution of one SRAM operation. We can observe from the figure that most of the SRAM operations need about 20 cycles to complete. Table 2: Latency Distribution for SRAM Access (cycles) Cycles of SRAM Access Percentage of SRAM Access

19

21

23

25

27

29

31

33

64.7

10.3

8.7

7.5

2.9

2.4

1.9

1.0

In Table 3, we compare the number of memory accesses required in our algorithm with that of a simple heap-based implementation. It takes only one SRAM access for a heap-based scheduler to find out the next packet to be sent. However, maintaining the heap requires O(logn) steps, and there are two heaps in the original DWCS implementation. Each step of heap reorganization requires six SRAM accesses. HILQ minimizes the number of SRAM accesses as compared to the heap-based DWCS implementation. The number of SRAM accesses per stream is close to a constant for HILQ, while it increases logarithmically with the number of active streams for the heapbased DWCS. As shown in Table 3, SRAM access is expensive. Combining results in Table 3, we can conclude that a simple heap-based DWCS would not scale as well as the HILQ for a large number of streams. Table 3: Memory Operation Comparison for Different Number of Streams No. of streams Memory access#

Heap HILQ

10 45.86 19.8

50 73.73 14.36

100 85.73 13.68

200 97.73 13.34

500 113.59 13.135

1000 125.59 13.068

2000 137.58 13.034

We test the scalability of the HILQ-based scheduler in simulation. There are two sets of tests, one without input packet traffic, and the other with input/output packet traffic. The objective of the first testing is to chart the actual scheduling overhead under different numbers of active streams. To obtain a large number of active streams, we use a separate microengine to generate a variety of pseudo-streams. In total, there are two microengines in use: one sends stream scheduling requests and the other runs the HILQ scheduler. Figure 6 shows the scheduling delay per active stream as we vary the number of active streams. It indicates that our HILQ-based DWCS scales well for a large number of active streams. As the number of streams increases, the average scheduling delay per active stream approaches 500 cycles/per stream. Readers may notice the high scheduling delay when the number of active streams is small (below 20). This is caused by a fixed amount of overhead in our HILQ scheduler. The scheduler has to read and check all indexes during a segment sweep even when there are only few active streams. Secondly, we want to show that the HILQ scheduler can actually produce high data throughput. Traffic is fed into the Developer Benchmark through a packet generator. We can adjust the packet speed and the number of streams through the packet generator and the configuration option in the simulator.

W. Shi et al.

Scheduling Delay per Stream (microenine cycles)

194

4000 3500 3000 2500 2000 1500 1000 500 0 0

500

1000

1500

2000

2500

No. of Active Streams per Segment in System Figure 6: Scalability - No. of Active Streams vs. Average Scheduling Delay

We linked our HILQ scheduler with microcode that performs packet receiving and transmission. There are two microengines, which run the receiver code that pulls packets from a Gigabit Ethernet port, and two other microengines run the transmission code that sends packets to another Gigabit Ethernet port. One more microengine is used for the HILQ scheduler. The receiver thread pulls each packet from the wire and puts it into the scheduler queue. The HILQ scheduler scans each segment linearly and sends out the waiting packets to an output queue. The transmission thread reads from the output queue and puts the packet on the wire. To sustain a certain data rate, the scheduler has to complete a packet scheduling cycle within a time limit. Figure 7 shows the throughput vs. the scheduling cycle for different sizes of packets. The figure indicates that large packets allow a higher time budget for a packet scheduler. When the packet size is 1K, a packet scheduler with about 900 cycle scheduling overhead is still capable of supporting a Gigabit data link. However, for small packet size like 128 bytes and 256 bytes, throughput starts to drop even with small scheduling overhead. 1200

Throughput (Mbps)

1000 800 600 400 128 256 512 1024

200 0

0

100

200

300

400

500

600

700

800

900 1000

Scheduling cycle

Figure 7: Throughput vs. Scheduling Cycles for Different Packet Sizes

One interesting result concerns the 512-byte packet size. There is about a 600 cycles time-budget for a packet scheduler to sustain a Gigabit link with a packet size of 512 bytes. This means that any packet scheduler that can schedule a packet in less than 600 cycles will not have a negative impact on the data rate. Since our HILQ has a limit of about 500 cycles, we anticipate that it should be capable of supporting Gigabit throughput for

Efficient Implementation of Packet Scheduling Algorithm

195

streams whose packet size are equal to or above 512 bytes. In Figure 8, the results indicate that with our HILQ implementation, the DWCS QoS scheduler is capable of supporting a Gigabit data link when the packet size is 512 bytes. For smaller size packets such as 256 bytes, the throughput is about 650MB/ps. This result is consistent with the scheduler time budget shown in Figure 7. For 512-byte packets, the scheduler is able to schedule about 260,000 packets per second. If each stream has only about 40 ñ 60 packets per second, our HILQ scheduler would be able to support about 4,000 ñ 6,000 streams. Note: Figures 6 and 7 refer to the number of active streams.

Throughput (Mbps)

1200 1000

800 600 400 200 0

512 byte packet256 byte packet-

0

100

200

300

400

500

600

Number of Active Streams

Figure 8: Throughput vs. No. of Active Streams for Different Sizes of Packets

The last result compares our DWCS approximation with the original DWCS. The single most important metric for evaluation is window constraint violation. Theoretically, DWCS does not have any loss tolerance violation when the bandwidth utilization factor is below 1.0 [2]. According to [2], the bandwidth utilization factor is computed as U = sum((1-Wi)Ci/Ti), i = 1,2,Ö, n, where Wi = xi/yi, and Ci is the amount of time required to serve a packet of stream i and Ti is the request period of stream i. However, when U exceeds 1.0, there will be window constraint violations. Results in Figure 9 show that our DWCS approximation does not create more window constraint violations than the heapbased implementation when U exceeds 1.0. The results are collected from a simulation having 200 streams. Each stream starts with a random Wi = xi/yi lying between 0.2 and 0.8. The range of x is 0 to 19 and the range of y is 0 to 29. Percentage of Window Constraint Violations

0.6 0.5 0.4 0.3 0.2 HILQ Heap

0.1 0

1

1.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

Utilization Factor

Figure 9: Comparison between HILQ and Heap Based Implementation under Different Utilization Factors

196

7

W. Shi et al.

Conclusion

This paper proposes a fast algorithm to implement a multimedia stream scheduler on the IXP Network Processor. A data structure called Hierarchically Indexed Linear Queue meets the requirements of Gigabit speeds by exploiting certain architectural attributes. High performance is gained by integrating send/receive and garbage collection threads. Our implementation can achieve O(1) packet insertion and selection time by hierarchically indexing a sparse queue and separation of the garbage collection threads. Although our design incurs a small amount of inaccuracy with regard to the standard DWCS specification, the inaccuracies are shown to be tolerable. Our performance results strongly support the benefits of HILQ by serving up to thousands of streams at Gigabit speeds.

8

References

[1] Richard West and Karsten Schwan, "Dynamic Window-Constrained Scheduling for Multimedia Applications," in Proceedings of the IEEE International Conference on Multimedia Computing and Systems (ICMCS), 1999. [2] Richard West and Christian Poellabauer, "Analysis of a Window-Constrained Scheduler for Real-Time and Best-Effort Packet Streams,î in Proceedings of the 21st IEEE Real-Time Systems Symposium (RTSS), 2000. [3] Richard West, Karsten Schwan and Christian Poellabauer, "Scalable Scheduling Support for Loss and Delay Constrained Media Streams", in Proceedings of the 5th IEEE Real-Time Technology and Applications Symposium (RTAS), 1999. [4] C. Liu and J. Layland. ìScheduling Algorithms for Multiprogramming in a Hard-RealTime Environment,î Journal of ACM, 20(1), January 1973. [5] P.Goyal, H.Vin, and H.Cheng, ìStart-time fair queueing: A scheduling algorithm for integrated services packet switching networks,î in IEEE SIGCOMMí96, IEEE,1996. [6] K. Parekh. ìA Generalized Processor Sharing Approach to Flow Control in Integrated Services Networks,î PhDthesis, Department of Electrical Engineering and Computer Science, MIT, February 1992. [7] ìIXP 1200 Network Processor: Software Reference Manualî, Part No. 278306-005. Sep. 2000. [8] ìIXP 1200 Network Processor: Programmerís Reference Manualî, Part No. 278304006. Sep. 2000. [9] ìIXP 1200 Network Processor: Development Tools Userís Guideî, Part No. 278302005. Oct. 2000. [10] P.Crowley, M.E.Fiuczynski, J.Baer, B.N.Bershad, ìCharacterizing processor architectures for Programmable Network Interfacess,î Proceedings of the 2000 International Conference on Supercomputing, pp.54-65, May. 2000. [11] Tammo Spalink, Scott Karlin, Larry Peterson, Yitzchak GottliebIn, ìBuilding a Robust Software-Based Router Using Network Processors,î Proceedings of the 18th ACM Symposium on Operating Systems Principles (SOSP'01) . pages 216--229, Chateau Lake Louise, Banff, Alberta, Canada, October 2001. [12] Xiaohu Qie, Andy Bavier, Larry Peterson, Scott Karlin ìScheduling Computations on a Software-Based Router,î Proceedings of SIMETRICS 2001, Jun 2001. [13] D.C. Stephen, J.C.R.Bennett, H.Zhang, ìImplementing scheduling algorithms in high-speed networksî, IEEE Selected Areas in Communications, pp.1145-1158, Vol 17, No.6, Jun.1999

Efficient Implementation of Packet Scheduling Algorithm

197

[14] A. Ioannou, M. Katevenis: "Pipelined Heap (Priority Queue) Management for Advanced Scheduling in High Speed Networks," Proc. IEEE Int. Conf. on Communications (ICC'2001), Helsinki, Finland, June 2001, pp. 2043-2047. [15] Jun Xu and Mukesh Singhal, "Cost-Effective Flow Table Designs for High-Speed Internet Routers: Architecture and Performance Evaluation," to appear in IEEE Transactions on Computers. [16] Ranjita Bhagwan, Bill Lin "Fast and Scalable Priority Queue Architecture for HighSpeed Network Switches," IEEE INFOCOMMí00, Tel Aviv, Vol. 2, pages 538-547, March 2000. [17] R. Brown, "Calendar queues: a fast 0(1) priority queue implementation for the simulation event set problem," Communications of the ACM, vol. 31, pp. 1220--1227, October 1988. [18] K. Bruce Erickson, Richard E. Ladner, Anthony LaMarca, ìOptimizing Static Calendar Queues,î IEEE Symposium on Foundations of Computer Science,1998. [19] Rajamar Krishnamurthy, Karsten Schwan, and Marcel Rosu, ìA Network CoProcessor-Based Approach to Scalable Media Streaming in Serversî, International Conference on Parallel Processing (ICPP), August 2000.

Distributed Dynamic Capacity Contracting: A Congestion Pricing Framework for Diff-Serv Murat Yuksel1 and Shivkumar Kalyanaraman2 1

2

CS Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180, USA [email protected] ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180, USA [email protected]

Abstract. In order to provide better Quality-of-Service (QoS) in large networks, several congestion pricing proposals have been made in the last decade. Usually, however, those proposals studied optimal strategies and did not focus on implementation issues. Our main contribution in this paper is to address implementation issues for congestion-sensitive pricing over a single domain of the differentiated-services (diff-serv) architecture of the Internet. We propose a new congestion-sensitive pricing framework Distributed Dynamic Capacity Contracting (Distributed-DCC), which is able to provide a range of fairness (e.g. max-min, proportional) in rate allocation by using pricing as a tool. Within the Distributed-DCC framework, we develop an Edge-to-Edge Pricing Scheme (EEP) and present simulation experiments of it.

1

Introduction

As multimedia applications with extensive traffic loads are becoming more common, better ways of managing network resources are necessary in order to provide sufficient QoS for those multimedia applications. Among several methods to improve QoS in multimedia networks and services, one particular method is to employ congestion pricing. The main idea is to increase service price when network congestion is more, and to decrease the price when congestion is less. Implementation of congestion pricing still remains a challenge, although several proposals have been made, e.g. [1,2]. Among many others, two major implementation obstacles can be defined: need for timely feedback to users about price, determination of congestion information in an efficient, low-overhead manner. The first problem, timely feedback, is relatively very hard to achieve in a large network such as the Internet. In [3], the authors showed that users do need feedback about charging of the network service (such as current price and prediction of service quality in near future). However, in our recent work [4], we illustrated that congestion control through pricing cannot be achieved if price changes are performed at a time-scale larger than roughly 40 round-trip-times (RTTs), which is not possible to implement for many cases. We believe that the The original version of this chapter was revised: The copyright line was incorrect.This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 

This work is sponsored by NSF under contract number ANI9819112, and cosponsored by Intel Corporation.

K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 198–210, 2002. c IFIP International Federation for Information Processing 2002

Distributed Dynamic Capacity Contracting

199

problem of timely feedback can be solved by placing intelligent intermediaries (i.e. software or hardware agents) between users and service providers. In this paper, we do not focus on this particular issue and leave development of such intelligent agents for future research. The second problem, congestion information, is also very hard to do in a way that does not need a major upgrade at network routers. However, in diff-serv [5], it is possible to determine congestion information via a good ingress-egress coordination. So, this flexible environment of diff-serv motivated us to develop a pricing scheme on it. In our previous work [6], we presented a simple congestion-sensitive pricing “framework”, Dynamic Capacity Contracting (DCC), for a single diff-serv domain. DCC treats each edge router as a station of a service provider or a station of coordinating set of service providers. Users (i.e. individuals or other service providers) make short-term contracts with the stations for network service. During the contracts, the station receives congestion information about the network core at a time-scale smaller than contracts. The station, then, uses that congestion information to update the service price at the beginning of each contract. Several pricing “schemes” can be implemented in that framework.

Fig. 1. DCC framework on diff-serv architecture. DCC models a short-term contract for a given traffic class as a function of price per unit traffic volume Pv , maximum volume Vmax (maximum number of bytes that can be sent during the contract) and the contract length T : Contract = f (Pv , Vmax , T )

(1)

Figure 1 illustrates the big picture of DCC framework. Customers can only access network core by making contracts with the provider stations placed at the edge routers. The stations offer contracts (i.e. Vmax and T ) to fellow users. Access to these available contracts can be done in different ways, what we call edge strategy. Two basic edge strategies are “bidding” (many users bids for an available contract) or “contracting” (users negotiate Pv with the provider for an available contract). So, edge strategy is the decision-making mechanism to identify which customer gets an available contract at the provider station. However, in DCC, we assumed that all the provider stations advertise the same price value for the contracts, which is very costly to implement over a wide

200

M. Yuksel and S. Kalyanaraman

area network. This is simply because the price value cannot be communicated to all stations at the beginning of each contract. In this paper, we relax this assumption by letting the stations to calculate the prices locally and advertise different prices than the other stations. We call this new version of DCC as Distributed-DCC. We introduce ways of managing the overall coordination of the stations for the common purposes of fairness and stability. We then develop a pricing scheme Edge-to-Edge Pricing (EEP). We illustrate stability of EEP by simulation experiments. We address fairness problems related to pricing, and show that the framework can achieve max-min and proportional fairness by tuning a parameter, called as fairness coefficient. The paper is organized as follows: In the next section, we position our work and briefly survey relevant work in the area. In Section 3 we describe DistributedDCC framework, and investigate various issues, such as price calculation, fairness, scalability. Next in Section 4, we develop the pricing scheme EEP. In Section 5, we make experimental comparative evaluation of EEP. We finalize with summary and discussions.

2

Related Work

There has been several pricing proposals, which can be classified in many ways such as static vs. dynamic, per-packet charging vs. per-contract charging. Although there are opponents to dynamic pricing in the area (e.g. [7,8]), most of the proposals have been for dynamic pricing (specifically congestion pricing) of networks. Examples of dynamic pricing proposals are MacKie-Mason and Varian’s Smart Market [1], Gupta et al.’s Priority Pricing [9], Kelly et al.’s Proportional Fair Pricing (PFP) [10], Semret et al.’s Market Pricing [11], and Wang and Schulzrinne’s Resource Negotiation and Pricing (RNAP) [12,2]. Odlyzko’s Paris Metro Pricing (PMP) [13] is an example of static pricing proposal. Clark’s Expected Capacity [14] and Cocchi et al.’s Edge Pricing [15] allow both static and dynamic pricing. In terms of charging granularity, Smart Market, Priority Pricing, PFP and Edge Pricing employ per-packet charging, whilst RNAP and Expected Capacity do not employ per-packet charging. Smart Market is based primarily on imposing per-packet congestion prices. Since Smart Market performs pricing on per-packet basis, it operates on the finest possible pricing granularity. This makes Smart Market capable of making ideal congestion pricing. However, Smart Market is not deployable because of its per-packet granularity and its many requirements from routers. In [16], we studied Smart Market and difficulties of its implementation in more detail. While Smart Market holds one extreme in terms of granularity, Expected Capacity holds the other extreme. Expected Capacity proposes to use long-term contracts, which can give more clear performance expectation, for statistical capacity allocation and pricing. Prices are updated at the beginning of each long-term contract, which incorporates little dynamism to prices. Our work, Distributed-DCC, is a middle-ground between Smart Market and Expected Capacity in terms of granularity. Distributed-DCC performs congestion pricing at

Distributed Dynamic Capacity Contracting

201

short-term contracts, which allows more dynamism in prices while keeping pricing overhead small. Another close work to ours is RNAP, which also mainly focused on implementation issues of congestion pricing on diff-serv. Although RNAP provides a complete picture for incorporation of admission control and congestion pricing, it has excessive implementation overhead since it requires all network routers to participate in determination of congestion prices. This requires upgrades to all routers similar to the case of Smart Market. Our work solves this problem by requiring upgrades only at edge routers rather than at all routers.

3

Distributed-DCC: The Framework

Distributed-DCC is specifically designed for diff-serv architecture, because the edge routers can perform complex operations which is essential to several requirements for implementation of congestion pricing. Each edge router is treated as a station of the provider. Each station advertises locally computed prices with information received from other stations. The main framework basically describes how to preserve coordination among the stations such that stability and fairness of the overall network is preserved. A Logical Pricing Server (LPS) (which can be implemented in a centralized or distributed manner, see Section 3.6) plays a crucial role. Figure 2 illustrates basic functions (which will be better understood in the following sub-sections) of LPS. The following sub-sections investigate several issues regarding the framework. 3.1

How to Calculate pij ?

Each ingress station i keeps a ”current” price vector pi , where pij is the price for flow from ingress i to egress j. So, how do we calculate the price-per-flow, pij ? The ingresses make estimation of budget for each edge-to-edge flow passing through themselves. Let ˆbij be the currently estimated budget for flow fij (i.e the flow from ingress i to egress j). The ingresses send estimated budgets to the corresponding egresses (i.e. ˆbij is sent from ingress i to egress j) at a deterministic time-scale. At the other side, the egresses receive budget estimations from all the ingresses, and also they make estimation of capacity cˆij for each particular flow. In other words, egress j calculates cˆij and is informed about ˆbij by ingress i. Egress j, then, penalizes or favors fij by updating its estimated budget value, i.e. bij = f (ˆbij , [parameters]) where [parameters] are optional parameters used for deciding whether to penalize or favor the flow. For example, if fij is passing through more congested areas than the other flows, egress j can penalize fij by reducing its budget estimation ˆbij . At another time-scale3 , egresses keep sending information to LPS. More specifically, for a diff-serv domain with n edge routers, egress j sends the following information to LPS: 3

Can be larger than ingress-egress time-scale, but should be less than contract length.

202

M. Yuksel and S. Kalyanaraman

1. the updated budget estimations of all flows passing through itself, i.e. bij for i = 1..n and i = j 2. the estimated capacities of all flows passing through itself, i.e. cˆij for i = 1..n and i = j

Fig. 2. Major functions of LPS. LPS receives information from egresses and, for each fij , calculates allowed capacity cij . Calculation of cij s is a complicated task which depends on bij s. In general, the flows should share capacity of the same bottleneck in proportion to their budgets. We will later define a generic algorithm to do capacity allocation task. LPS, then, sends the following information to ingress i:   1. the total estimated network capacity C (i.e. C = i j cˆij ) 2. the allowed capacities to each edge-to-edge flow starting from ingress i, i.e. cij for j = 1..n and j = i Now, the pricing scheme at ingress i can calculate price for each flow by using cij and ˆbij . An example pricing scheme will be described in Section 4. 3.2

Budget Estimation at Ingresses

In order to determine user’s real budget The ingress stations perform very trivial operation to estimate budgets of each flow, ˆbij . The ingress i basically knows its current price for each flow, pij . When it receives a packet it just needs to determine which egress station the packet is going to. Given that the ingress station has the addresses of all egress stations of the same diff-serv domain, it can find out which egress the packet is going to. So, by monitoring the packets transmitted for each flow, the ingress can estimate the budget of each flow. Let xij be the total number of packets transmitted for fij in unit time, then the budget estimate for fij is ˆbij = xij pij . Notice that this operation must be done at the ingress rather than egress, because some of the packets might be dropped before arriving at the egress. This causes xij to be measured less, and hence causes ˆbij to be less than it is supposed to be.

Distributed Dynamic Capacity Contracting

3.3

203

Congestion-Based Capacity Estimation at Egresses

The essence of capacity estimation in Distributed-DCC is to decrease the capacity estimation when there is congestion indication(s) and to increase it when there is no congestion indication. This will make the prices congestion-sensitive, since the pricing scheme is going to adjust the price according to available capacity. In this sense, several capacity estimation algorithms can be used, e.g. Additive Increase Additive Decrease (AIAD), Additive Increase Multiplicative Decrease (AIMD). We now provide a full description of such an algorithm. With a simple congestion detection mechanism (such as marking of packets at interior routers when congested), egress stations make a congestion-based estimation of the capacity for the flows passing through themselves. Egress stations divide time into deterministic observation intervals and identify each observation interval as congested or non-congested. Basically, an observation interval is congested if a congestion indication was received during that observation interval. At the end of each observation interval, the egresses update the estimated capacity. Then, egress j calculates the estimated capacity for fij at the end of observation interval t as follows:  β ∗ µij (t), congested cˆij (t) = c, non-congested cˆij (t − 1) + ˆ where β is in (0,1), µij (t) is the measured output rate of fij during observation interval t, and ˆ c is a pre-defined increase parameter. This algorithm is a variant of well-known AIMD. 3.4

Capacity Allocation to Edge-to-Edge Flows

LPS is supposed to allocate the total estimated network capacity C to edge-toedge flows in such a way that the flows passing through the same bottleneck should share the bottleneck capacity in proportion to their budgets, and also the flows that are not competing with other flows should get all the available capacity on their route. The complicated issue is to do this without knowledge of the topology for network core. We now propose a simple and generic algorithm to perform this centralized rate allocation within Distributed-DCC framework. First, at LPS, we introduce a new information about each edge-to-edge flow fij . A flow fij is congested, if egress j has been receiving congestion indications from that flow recently (we will later define what “recent” is). At LPS, let Kij determine whether fij is congested or not. If Kij > 0, LPS determines fij as congested. If not, it determines fij as non-congested. Let’s call the time-scale at which LPS and egresses communicate as LPS interval. At every LPS interval t, LPS calculates Kij as follows:  ˆ fij was congested at t − 1 k, Kij (t) = (2) max(0, Kij (t − 1) − 1), fij was non-congested at t − 1 where kˆ is a positive integer. Notice that kˆ parameter defines how long a flow will stay in “congested” state after the last congestion indication. So, kˆ defines

204

M. Yuksel and S. Kalyanaraman

the time-line to determine if a congestion indication is “recent” or not. Note that instead of setting Kij to kˆ at every congestion indication, several different methods can be used for this purpose, but we proceed with the method in (2). Given the above method to determine whether a flow is congested or not, we now describe the algorithm to allocate capacity to the flows. Let F be the set of all edge-to-edge flows in the diff-serv domain, and Fc be the set of congested edge-to-edge flows. Let Cc be the accumulation of cˆij s where fij ∈ Fc . Further, let Bc be the accumulation of bij s where fij ∈ Fc . Then, LPS calculates the allowed capacity for fij as follows:  bij C ,K >0 cij = Bc c ij cˆij , otherwise The intuition is that if a flow is congested, then it must be competing with other congested flows. So, a congested flow is allowed a capacity in proportion to its budget relative to budgets of all congested flows. Since we assume no knowledge about the interior topology, we approximate the situation by considering these congested flows as if they are traversing a single bottleneck. If knowledge about the interior topology is provided, one can easily develop better algorithms by sub-grouping the congested flows that are passing through the same bottleneck. If a flow is not congested, then it is allowed to use its own estimated capacity, which will give enough freedom to utilize capacity available to that flow. 3.5

Fairness

We examine the issues regarding fairness in two main cases: – Single-bottleneck case: The pricing protocol should charge same price to users of same bottleneck. In this way, among users of same bottleneck, the ones with more budget will be given more capacity. The intuition behind this reasoning is that the cost of providing capacity to each customer is the same. – Multi-bottleneck case: The pricing protocol should charge more to users whose traffic traverses more bottlenecks and causes more costs. So, other than proportionality to user budgets, we also want to allocate less capacity to users whose flows are traversing more bottlenecks than the others. For multi-bottleneck networks, two main types of fairness have been defined: max-min, proportional [10]. In max-min fairness, flows get equal share of bottlenecks, while in proportional fairness flows get penalized according to number of bottlenecks they traverse. Depending on cost structure and user utilities, provider may want to choose max-min or proportional fairness. So, we would like to have ability of tuning pricing protocol such that fairness of its rate allocation is in the way provider wants. To achieve the above fairness objectives in Distributed-DCC, we introduce new parameters for tuning rate allocation to flows. In order to penalize fij , egress j can reduce ˆbij , which causes decrease in cij . It uses following function: bij = f (ˆbij , r(t), α, rmin ) =

ˆbij rmin + (rij (t) − rmin ) ∗ α

Distributed Dynamic Capacity Contracting

205

where rij (t) is the estimated congestion cost caused by fij , rmin is the minimum possible congestion cost, and α is fairness coefficient. Instead of ˆbij , the egress j now sends bij to LPS. When α is 0, Distributed-DCC is employing max-min fairness. As it gets larger, the flow gets penalized more and rate allocation gets closer to proportional fairness. However, if it is too large, then the rate allocation will get away from proportional fairness. Let α∗ be the α value where the rate allocation is proportionally fair. If the estimation rij (t) is accurate, then α∗ = 1. Assuming that severity of each bottleneck is the same, we can directly use the number of bottlenecks fij is traversing in order to calculate rij (t) and rmin . In such a case, rmin will be 1 and rij (t) should be number of bottlenecks the flow is passing through. If the interior nodes increment a header field of the packets at the time of congestion, then the egress station can estimate the number of bottlenecks the flow is traversing. We skip description of such an estimation algorithm to keep reader’s focus on major issues. 3.6

Scalability

There are mainly two issues regarding scalability: LPS, the number of flows. First of all, the flows are not per-connection basis, i.e. all the traffic going from edge router i to j is counted as only one flow. This actually relieves the scaling of operations happening on per-flow basis. The number of flows in the system will be n(n − 1) where n is the number of edge routers in the diff-serv domain. So, indeed, scalability of the flows is not a problem for the current Internet since number of edge routers for a single diff-serv domain is very small. LPS can be scaled in two ways. First idea is to implement LPS in a fully distributed manner. Edge stations exchange information with each other (like link-state routing). Basically, each station sends total of n − 1 messages to other stations. So, this will increase overhead on network because of the extra messages, i.e. complexity will increase from O(n) to O(n2 ) in terms of number of messages. Alternatively, LPS can be divided into multiple local LPSs which synchronize among themselves to maintain consistency. This way the complexity of number of messages will reduce. However, this will be at a cost of some optimality again.

4

Edge-to-Edge Pricing Scheme (EEP)

For flow fij , Distributed-DCC framework provides an allowed capacity cij and an estimation of total user budget ˆbij at ingress i. So, the provider station at ingress i can use these two information to calculate price. We propose a simple price formula to balance supply cij and demand ˆbij : pij =

ˆbij cij

(3)

Also, the ingress i uses the total estimated network capacity C in calculating the Vmax contract parameter defined in (1). A simple method for calculating

206

M. Yuksel and S. Kalyanaraman

Vmax is Vmax = C ∗T where T is the contract length. This allows all the available capacity to be contracted by a single flow, which is a loose admission control. More conservative admission control should be used to calculate Vmax . We now prove optimality of (3) for a single-bottleneck network. We skip the proof for a multi-bottleneck network for space considerations. We model user i’s utility with the well-known 4 [10] function ui (x) = wi log(x), where x is bandwidth given to the user and wi is user i’s budget (i.e. willingness-to-pay). Suppose pi is the price advertised to user i. Then, user i will maximize his surplus Si by contracting for xi = wi /pi . So, the provider can now figure out what price to advertise to each user by maximizing the social welfare W = S + R, where R is the provider revenue. Let K(x) = kx be a linear function and be the cost of providing x amount of capacity to a user, where k is a positive constant. Then social welfare will be: W =

n 

[ui (xi ) − kxi ]

i=1

 We maximize W subject to i xi ≤ C, where C is the total available capacity. To maximize W , all the available capacity must be given to users since they have strictly increasing utility. Lagrangian and its solution for that system will be: W

=

n 

n  ui (xi ) − kxi + λ( xi − C)

i=1

xj

wj = n i=1

i=1

wi

C, j = 1..n

(4)

This result shows that welfare maximization of the described system can be done by allocating capacity to the users proportional to their budget, wi , relative to total user budget. Since the user will contract for xi = wi /pi when advertised a price of pi , then the optimum price for provider to advertise (i.e. p∗ ) can be calculated as follows: n wi ∗ p = pi = i=1 C i.e. ratio of total budget to available capacity. So, provider should charge same price to users of same bottleneck route. EEP does that for an edge-to-edge route.

5

Simulation Experiments and Results

We present ns [17] simulation experiments of EEP on single-bottleneck and multi-bottleneck topology. Our goals are to illustrate fairness and stability properties of the framework. The single-bottleneck topology is shown in Figure 3-a and the multi-bottleneck topology is shown in Figure 3-b. The white nodes are edge nodes and the gray nodes are interior nodes. To ease understanding 4

Wang and Schulzrinne introduced a more complex version in [12].

Distributed Dynamic Capacity Contracting

207

the experiments, each user sends its traffic to a separate egress. For the multibottleneck topology, one user sends through all the bottlenecks (i.e. long flow), crossed by the others. Bottleneck links have a capacity of 10Mb/s and all other links have 15Mb/s. Propagation delay on each link is 5ms, and users send UDP traffic with an average packet size of 1000B. Interior nodes mark the packets when local queue exceeds 30 packets. In the multi-bottleneck topology they increment a header field instead of just marking. Buffer size is assumed infinite.

(a) Single-bottleneck network

(b) Multi-bottleneck network

Fig. 3. Topologies for Distributed-DCC experiments. Each user flow maximizes its surplus by contracting for b/p amount of capacity, where b is its budget and p is price. b is randomized according to truncatedNormal distribution with mean b. We will refer to this mean b as flow’s budget. Ingresses send budget estimations to egresses at every observation interval. LPS sends information to ingresses at every LPS interval. Contracting takes place at every 4s, observation interval is 0.8s, and LPS interval is 0.16s. The parameter kˆ is set to 25 (see Section 3.4). ˆ c is set to 1 packet (i.e. 1000B), the initial value of cˆij for each flow fij is set to 0.1Mb/s, and β is set to 0.95. 5.1

Experiment on Single-bottleneck Topology

We run a simulation experiment for EEP on the single-bottleneck topology, which is represented in Figure 3-a. There are 3 users with budgets of 10, 20, 30 respectively for users 1, 2, 3. Total simulation time is 15000s. Initially, only user 1 is active and after every 5000s one of the other users gets active. Figure 4-a shows the flow rates averaged over 200 contract periods. We see the flows are sharing the bottleneck capacity almost in proportion to their budgets. The distortion in rate allocation is caused because of the assumptions that the generic edge-to-edge capacity allocation algorithm makes (see Section 3.4). Also, Figure 4-b shows the average prices charged to flows over 200 contract periods. As new users join in, EEP increases the price for balancing supply and demand. Figure 4-c shows the bottleneck queue size. Notice that queue sizes make peaks transiently at the times when new users gets active. Otherwise, the queue size is controlled reasonably and the system is stable. The reason behind the transient peaks is that the parameter Vmax is not restricted which causes the newly joining flow to contract for a lot more than the available capacity. Also, average utilization of the bottleneck link was more than 90%.

208

5.2

M. Yuksel and S. Kalyanaraman

Experiments on Multi-bottleneck Topology

On a multi-bottleneck network, we would like illustrate two properties: – Property 1: provision of various fairness in rate allocation by changing the fairness coefficient α (see Section 3.5) – Property 2: adaptiveness of capacity allocation algorithm (see Section 3.4) In order to illustrate Property 1, we run a series of experiments for EEP with different α values. We use a larger version of the topology represented in Figure 3-b. In the multi-bottleneck topology there are 10 users and 9 bottleneck links. Total simulation time is 10,000s. Initially, the long flow is active. After each 1000s, one of the other cross flows gets active. So, as the time passes the number of bottlenecks in the system increases. We are interested in the long flow’s rate, since it is the one that cause more congestion costs than the other flows. Figure 4-d shows average rate of the long flow versus number of bottlenecks in the system. As expected, the long flow gets lesser capacity as α increases. When α = 0, rate allocation to flows is max-min fair. Observe that when α = 1, rate allocation follows the proportionally fair rate allocation. This variation in fairness is basically achieved by advertisement of different prices to the flows. Figure 4-e shows the average price that is advertised to the long flow as number of bottlenecks in the system increases. We can see that the price advertised to the long flow increases as number of bottlenecks increases. As α increases, framework becomes more responsive to the long flow by increasing its price more sharply. Finally, to illustrate Property 2, we ran an experiment on the topology in Figure 3-b with small changes. We increased capacity of the bottleneck at node D from 10 Mb/s to 15Mb/s. Initially, all the flows have equal budget of 10 units. Total simulation time is 30000s. Between times 10000 and 20000, budget of flow 1 is temporarily increased to 20 units. α is set to 0. All the other parameters are exactly the same as in the single-bottleneck experiments of the previous section. Figure 4-f shows the given volumes averaged over 200 contracting periods. Until time 10000s, flows 0, 1, and 2 share the bottleneck capacities equally presenting a max-min fair allocation because α is 0. However, flow 3’s rate is larger, because bottleneck node D has extra capacity. This is achieved by the freedom given to individual flows by the capacity allocation algorithm (see Section 3.4). Between times 10000 and 20000, flow 2 gets a step increase in its rate because of the step increase in its budget. In result of this, flow 0 gets a step decrease in its volume. Also, flows 2 and 3 adapt themselves to the new situation, trying to utilize the extra capacity leftover from the reduction in flow 0’s rate. So, flows 2 and 3 get a step decrease in their rates. After time 20000, flows restore to their original rates, illustrating adaptiveness of the framework.

6

Summary

In this paper, we presented a new framework, Distributed-DCC, for congestion pricing in a single diff-serv domain. Distributed-DCC can provide a contracting

Distributed Dynamic Capacity Contracting

209

framework based on short-term contracts between multimedia (or any other elastic, adaptive) application and the service provider. Contracting improves QoS if appropriate admission control and provisioning techniques are used. In this paper, we focused on pricing issues. 10

8

Averaged Price of Flows ($/Mb)

Averaged Rates of Flows (Mb/s)

5

Flow 0 Flow 1 Flow 2

9

7 6 5 4 3 2

Flow 0 Flow 1 Flow 2

4

3

2

1

1 0

0

2000

4000

6000 8000 Time (seconds)

10000

12000

0

14000

0

2000

4000

(a) Averaged flow rates

6000 8000 Time (seconds)

9

1200

7

1000 800 600 400

6 5 4 3

1 0

2000

4000

6000 8000 Time (seconds)

10000

12000

0

14000

0

1

(c) Bottleneck queue length

4

3 2.5

3 4 5 Number of Bottlenecks

12

alpha = 0.0 alpha = 0.25 alpha = 0.50 alpha = 0.75 alpha = 1.00 alpha = 1.25 alpha = 1.50 alpha = 1.75 alpha = 2.00 alpha = 2.25 alpha = 2.50

3.5

2

6

7

8

(d) Rate of long flow Flow 0 Flow 1 Flow 2 Flow 3

10 Averaged Rates of Flows (Mb/s)

Price Advertised to the Long Flow ($/Mbits)

5

14000

2

200

4.5

12000

alpha = 0.0 alpha = 0.25 alpha = 0.50 alpha = 0.75 alpha = 1.00 alpha = 1.25 alpha = 1.50 alpha = 1.75 alpha = 2.00 alpha = 2.25 alpha = 2.50 Proportional fair rate

8

Long Flow Rate(Mb/s)

Bottleneck Queue Size (packets)

1400

0

10000

(b) Prices to flows

2 1.5 1

8

6

4

2

0.5 0

0

1

2

3 4 5 Number of Bottlenecks

6

(e) Price to the long flow

7

8

0

0

5000

10000

15000 Time (seconds)

20000

25000

30000

(f) Averaged flow rates

Fig. 4. (a)-(c): Results of single-bottleneck experiment for EEP. (d)-(f): Results of EEP experiments on multi-bottleneck topology. Main contribution of the paper is to develop an easy-to-implement congestion pricing architecture which provides flexibility in rate allocation. We investigated fairness issues within Distributed-DCC and illustrated ways of achieving a range of fairness types (i.e. from max-min to proportional) through congestion pricing

210

M. Yuksel and S. Kalyanaraman

under certain conditions. The fact that it is possible to achieve various fairness types within a single framework is very encouraging. We also developed a pricing scheme, EEP, within the Distributed-DCC framework, and presented several simulation experiments. Future work should include investigation of issues related to Distributed-DCC on multiple diff-serv domains. Also, the framework should be supported by admission control techniques which will tune the contract parameter Vmax and address minimum QoS settings in SLAs.

References 1. J. K. MacKie-Mason and H. R. Varian, Pricing the Internet, Kahin, Brian and Keller, James, 1993. 198, 200 2. X. Wang and H. Schulzrinne, “An integrated resource negotiation, pricing, and QoS adaptation framework for multimedia applications,” IEEE Journal of Selected Areas in Communications, vol. 18, 2000. 198, 200 3. A. Bouch and M. A. Sasse, “Why value is everything?: A user-centered approach to Internet quality of service and pricing,” in Proceedings of IWQoS, 2001. 198 4. M. Yuksel, S. Kalyanaraman, and B. Sikdar, “Effect of pricing intervals on congestion-sensitivity of network service prices,” Tech. Rep. ECSE-NET-2002-1, Rensselaer Polytechnic Institute, ECSE Networks Lab, 2002. 198 5. S. Blake et. al, “An architecture for Differentiated Services,” RFC 2475, 1998. 199 6. R. Singh, M. Yuksel, S. Kalyanaraman, and T. Ravichandran, “A comparative evaluation of Internet pricing models: Smart market and dynamic capacity contracting,” in Proceedings of Workshop on Information Technologies and Systems (WITS), 2000. 199 7. A. M. Odlyzko, “Internet pricing and history of communications,” Tech. Rep., AT & T Labs, 2000. 200 8. I. Ch. Paschalidis and J. N. Tsitsiklis, “Congestion-dependent pricing of network services,” IEEE/ACM Transactions on Networking, vol. 8, no. 2, 2000. 200 9. A. Gupta, D. O. Stahl, and A. B. Whinston, Priority pricing of Integrated Services networks, Eds McKnight and Bailey, MIT Press, 1997. 200 10. F. P. Kelly, A. K. Maulloo, and D. K. H. Tan, “Rate control in communication networks: Shadow prices, proportional fairness and stability,” Journal of Operations Research Society, vol. 49, pp. 237–252, 1998. 200, 204, 206 11. N. Semret, R. R.-F. Liao, A. T. Campbell, and A. A. Lazar, “Market pricing of differentiated Internet services,” in Proceedings of IWQoS, 1999, pp. 184–193. 200 12. X. Wang and H. Schulzrinne, “Pricing network resources for adaptive applications in a Differentiated Services network,” in Proceedings of INFOCOM, 2001. 200, 206 13. A. M. Odlyzko, “A modest proposal for preventing Internet congestion,” Tech. Rep., AT & T Labs, 1997. 200 14. D. Clark, Internet cost allocation and pricing, Eds McKnight and Bailey, MIT Press, 1997. 200 15. R. Cocchi, S. Shenker, D. Estrin, and L. Zhang, “Pricing in computer networks: Motivation, formulation and example,” IEEE/ACM Transactions on Networking, vol. 1, December 1993. 200 16. M. Yuksel and S. Kalyanaraman, “Simulating the Smart Market pricing scheme on Differentiated Services architecture,” in Proceedings of CNDS part of SCS Western Multi-Conference, 2001. 200 17. “UCB/LBLN/VINT network simulator - ns (version 2),” http://www-mash.cs.berkeley.edu/ns, 1997. 206

Fair Stateless Aggregate Traffic Marking Using Active Queue Management Techniques A. Das, D. Dutta, and A. Helmy University of Southern California [email protected], [email protected], [email protected]

Abstract. The differentiated services architecture promises to provide QoS guarantees through scalable service differentiation among multimedia and best-effort flows in the Internet. Traffic marking is an important component of this Diffserv framework. In this paper, we propose two new stateless, scalable and fair aggregate markers for TCP aggregates and UDP multimedia aggregates. We leverage stateless Active Queue Management (AQM) algorithms to design markers that ensure fair and efficient token distribution among individual flows of an aggregate. We present the Probabilistic Aggregate Marker (PAM), that uses the token bucket burst size to probabilistically mark incoming packets to achieve TCP-friendly and fair marking, and the Stateless Aggregate Fair Marker (F-SAM) that approximates fair queueing techniques to isolate flows while marking packets of the aggregate. Our simulation results show that our marking strategies show upto 30% improvement over other commonly used markers while marking flow aggregates. When applied to aggregate TCP flows consisting of long-lived flows(elephants) and short lived web flows(mice), our F-SAM marker prevents any bias against short flows and helps the mice to win the war against elephants.

1

Introduction

The Differentiated Service architecture (Diffserv) [2] [16] aims to provide statistical QoS guarantees within IP. Unlike Intserv [22], Diffserv is stateless, and, hence, scalable. It uses fine grained, per-flow marking at the network edges to classify flows into the various classes, and, then, applies coarser per-class service differentiation (Per-Hop Behavior) at the network core. There are two main types of Per-hop Behaviors(PHB) currently defined, the Expedited Forwarding(EF) PHB, and the Assured Forwarding(AF) PHB. In this paper, we design packet markers for an Assured Forwarding PHB domain [13]. Traffic is marked at the edge into different drop-priority classes, according to a traffic profile based on service level agreements (SLA) between two domains. The network core uses simple AQM techniques to provide preferential packet dropping among the classes. Since the SLA at an edge domain applies to the total egress traffic, we need to perform traffic marking on the aggregated traffic at the edge. Simple token-bucket based markers however cannot distinguish between flows within an aggregate. In particular, misbehaving, The original version of this chapter was revised: The copyright line was incorrect.This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 211–223, 2002. c IFIP International Federation for Information Processing 2002

212

A. Das, D. Dutta, and A. Helmy

non-responsive flows can get more tokens. Also there is no provision for preventing long TCP flows from squeezing out short duration web traffic. Token-bucket markers aggravate bursty behavior, encourage misbehaving flows, and increase packet drops; this can severely affect both TCP and multimedia performance. Thus there is a need to mark a traffic aggregate fairly and efficiently according to a given traffic specification. We need to ensure that we protect the TCP traffic in an aggregate from misbehaving flows. Also, we need to ensure that within an aggregate of multimedia UDP flows, the congestion-controlled well-behaved flows do not suffer at the cost of the other non congestion-controlled UDP flows. Again, if we consider just TCP traffic, we do not want the short term TCP flows (mice) to lose out to the long lived TCP flows (elephants). Hence we need to look at traffic markers which can provide fairness among any aggregate without maintaining per-flow state, thus ensuring higher scalability. Traffic belonging to a single aggregate is marked into conformant (IN) and non-conformant (OUT) packets. The issues in Diffserv traffic marking are also similar to those in Active Queue Management [7] schemes. Well established work in active queue management such as RED [9], CHOKE [17] and CSFQ [18] provide certain degrees of fairness to flows and/or better throughput for flows by probabilistic dropping of packets from the queue, without maintaining per-flow state. We leverage AQM techniques to design two novel stateless, fair, aggregate markers. We show through extensive simulations that our methods achieve fairness in token allocation between flows without maintaining per-flow state in a variety of scenarios containing UDP, TCP and mixed traffic. We are not aware of any other work that utilizes AQM techniques to mark aggregate traffic in a fair and stateless fashion. Our techniques hold promise to control well behaved multimedia traffic and TCP traffic from misbehaving traffic within the Diffserv framework. The rest of this document is outlined as follows. Section 2 presents related work on markers. A brief background on traffic marking is given in Section 3 and our first scheme, PAM, is describes in Section 4. Section 5 details our second scheme, F-SAM. Our simulation results are discussed in Section 6. Section 7 outlines future work and concludes.

2

Related Work

Packet marking is a well visited topic. Many current markers perform sophisticated per-flow marking [1] [21]. Others mark packet aggregates without maintaining any per-flow states. Stateless markers are simpler to implement, and, hence, desirable. There has been some work in stateless, aggregate marking. They typically rely on metering the average rates and comparing them against a token traffic specification [11], [12]. However these markers do not address fairness issues within an aggregate. A sophisticated aggregated marker can be found in Fang. et. al. [8]. It does probabilistic marking based on the average aggregate rates, and in this respect, is similar to PAM. However instead of a RED-like transfer function based on the

Fair Stateless Aggregate Traffic Marking

213

average token-burst size, as in our case, they rely on the average incoming rate. Thus, their marker cannot handle bursty flows or the bias against short lived TCP flows while marking in a congested network. Maintaining per-flow state can solve the problem of fair aggregate traffic marking. However, these schemes are not scalable when the number of flows entering a marker is varying and dynamic. Yeom and Reddy [21] address fair aggregate marking issues, but they assume that the individual flows within the aggregate have already been pre-allocated individual contract rate from the aggregate marker tokens. Also their algorithm is per flow based, and entails calculating individual flow throughput information using a TCP throughput model which requires Round Trip Time and packet size information for each flow. Unlike our markers, they only handle TCP flows. In [19], the authors look specifically at the issue of sharing excess network bandwidth among traffic aggregates. However, they do not fairly marking individual flows within an aggregate, while we do. So their aggregate marker looks at inter-aggregate fairness as opposed to intra-aggregate fairness, as in our case. In [1] the authors give preferential treatment to short-lived TCP flows in an aggregate, and perform a max-min fair allocation (of tokens) on the remaining TCP flows. But they assume that their marker has state about all the component TCP flows in the aggregate, which might not be practical or scalable. Also, their scheme works only for TCP. In [14], Lloyd et. al. use the FRED algorithm to create a fair aggregate traffic marker. However this requires per-active flow computation at the marker, and is also more complex to implement than either of our markers. Also, they do not address the issues in fairness between long-lived and short-lived TCP flows. Protection of TCP traffic and well-behaved multimedia traffic from each other and from misbehaving UDP traffic has been a great concern. In [3], the authors define thresholds for different classes of traffic. They use a dynamic threshold based scheme on top of RED to provide inter-class and intra class fairness between TCP and well behaved tagged multimedia traffic. Thus, they ensure that the well behaved UDP flows are not punished just due to their non-TCP friendly properties. However, their scheme does not use the Diffserv framework. Within a Diffserv framework, our markers will achieve the same goal if we mark TCP and UDP traffic with separate marker instances and decide to put the TCP and the UDP flows in different Diffserv classes. Also, unlike them, we present a class of fair markers that are based on AQM techniques. We are not aware of any previous work in traffic marking that solves the problem of fair token distribution among flows within an aggregate without maintaining per-flow state, within the Diffserv framework. Besides, we are not aware of any marking techniques based on AQM techniques such as RED [9] and CSFQ [18].

214

3

A. Das, D. Dutta, and A. Helmy

Traffic Marking

Traffic marking is an essential component of the Diffserv architecture. The traffic marker looks at the incoming packets and compares it with a given traffic profile (for example, a token bucket(TB) characterization of a SLA between two domains). In-profile(IN) packets are marked with an identifier that indicates a higher priority. If packets are out of profile (OUT), it is marked with a lower priority. The main challenge in marking is to efficiently and fairly distribute the allocated IN tokens (specified by the SLA) among individual flows of the edge network entering the marker. One solution is to maintain states of all flows and allocate IN-tokens according to the SLA. Clearly this is not scalable. The other alternative is to mark aggregates using a stateless algorithm. This is our approach. The problem of aggregate marking is quite similar to the issue of queue management and scheduling. The queue receives packets from various flows and has to decide which packets from the incoming aggregate to buffer, which to drop, and how to allocated bandwidth fairly among flows. We can view a token bucket specification as a queue with the token bucket (TB) burst size as the queue size and the token arrival rate as the queue’s link bandwidth. Marking an arriving packet as IN is same as queueing a packet for transmission, and marking it as OUT is equivalent to dropping an incoming packet from the queue. Both the average queue size and the average token bucket size give an indication of the congestion level at the queue/TB. However a small average queue size is equivalent to a large average token bucket size and vice versa. The problem of fair token distribution at a marker (using a Token Bucket traffic specification) among packets of various incoming flows, is equivalent to efficient buffer management and scheduling of incoming packets at a queue in a fair manner. Active Queue Management techniques such as Random Early Detection(RED) [9] pro-actively drop packets before a queue overflow event, and allow for higher throughput and better fairness by reducing synchronization effects, bursty packet losses and biases against low bandwidth-bursty TCP flows. In addition, queue management techniques like Core Stateless Fair Queueing(CSFQ) [18], Flow Random Early Detection(FRED) [15], CHOKE [17] and Fair Queueing [6] try to distribute link bandwidth at the queue fairly among all flows of the aggregate. We apply two algorithms from sample AQM techniques - CSFQ and RED, to aggregate Diffserv packet marking. We observe that the use of these techniques yields superior marker performance, in terms of throughput and fairness. In the next two section, we introduce our two stateless AQM based marking schemes. The first one, PAM, is based on RED and the second one, F-SAM, is based on CSFQ.

4

Probabilistic Aggregate Marker

Our Probabilistic Aggregate Marker (PAM) is based on RED (Random Early Detection) [9]. The aggregate traffic tries to consume tokens from the token

Fair Stateless Aggregate Traffic Marking

215

bucket at a certain rate. We maintain an exponentially weighed moving average (EWMA) of the number of tokens in the token bucket. On every incoming packet, we look at this average token bucket size. If this bucket size falls below a certain threshold minth , all packets are to be marked with a lower priority. If the bucket size varies between minth and maxth , we mark the packet as OUT based on a probability function that depends on the size of the token bucket. If the token bucket size exceeds maxth , we mark the packet as IN. More formally, our probability function of marking as OUT P (x) can be written as follows. P (x) =

  

Pmax −Pmin maxth −minth

1 × (maxth − x) 0

: : :

x < minth minth < x < maxth x > maxth

(1)

where x is the average size of the token bucket. Our probability function for marking a packet as OUT, therefore a modification of the RED probability function for accepting a packet into the queue. This marking scheme ensures that the ratio of IN packets to OUT packets is flow-independent. Hence, the flow that pushes more traffic into the edge will, on the average, have a greater number of OUT packets; hence it will have more of its packets will be dropped if the core is congested. We show in Section 6 show that this scheme will be fairer to TCP flows in the presence of a misbehaving non-responsive flow, than simple token bucket schemes. The main advantages of PAM include its simplicity in implementation and deployment, and its stateless property.

5

Fair Stateless Aggregate Marker

In this section, we propose a very efficient, stateless aggregate fair marker FSAM that employs algorithms from Core Stateless Fair Queueing (CSFQ) [18] like algorithm. CSFQ provides approximate max-min fairness while buffering incoming packets at a queue, by using a probabilistic dropping function based on the average rate of the flow to which the packet belongs. This rate information, instead of being calculated at the queue using per-flow techniques is calculated near the source of the flow and inserted in every packet header. Similarly, we maintain the rate information of a flow in the packet header itself, and use this to calculate a token probability, on a per-packet basis in the edge marker. Packets which can get through the probabilistic dropping function based on their rate information to get tokens from the token bucket are marked as IN, while others are marked as OUT. The queue, the output link speed, and drop probability in [18] are replaced by the token bucket, token bucket rate and (1 − P [in token]). In F-SAM, the rate information in each packet header is calculated and filled by the ingress node when the flow enters the edge domain. Since each ingress node is responsible for maintaining the rate of only the flows originating from it, this per-flow rate calculation is simple and scalable. At the egress edge router, the edge marker needs to calculate the fair rate, f , allocated to the flows, and then calculates the token allocation probability of a packet as min(1, fr ), where r is the rate of the corresponding flow. So the expected token allocation rate to

216

A. Das, D. Dutta, and A. Helmy

a packet belonging to a flow of rate r, is then min(r, f ), which corresponds to the max-min fair rate of fair-queuing. As in [18], the fair-rate f , is calculated by an iterative method based on the average aggregate arrival rate of packets A, aggregate token allocation rate of packets R‘, and the token bucket rate C. Note that the fair-rate calculation algorithm does not need any per-flow measurements. Due to lack of space, we do not go into the specific algorithm details (see [18] or [4] for more details.) Thus, F-SAM can distribute tokens among various flows in an approximate max-min fair manner like CSFQ. Note that fair token distribution among flow aggregates does not translate into a exactly fair division of throughput at the network core among TCP and UDP flows flows, if the core only implements simple RIO queueing. This is because the loss of any packet of a TCP flow would result in its halving its sending rate, while it does not affect the sending rate of the UDP flow. However, preventing large UDP flows from getting more IN tokens than TCP flows would significantly increase the throughput obtained by the TCP flows in presence of misbehaving traffic. Also, since this is a probabilistic marker, it will be not encourage bursty losses. Thus, even in the case of all-TCP aggregate traffic, the throughput and fairness obtained by individual TCP flows will be much higher than simple token bucket markers. Note that even in case of aggregates of multimedia UDP flows, F-SAM will ensure fair token distribution to all the individual flows within the aggregate. Thus, given separate traffic specifications for UDP and TCP aggregates at a diffserv marker node, one could have two instances of the F-SAM marker operating on TCP and UDP aggregates separately, while maintaining fair token allocation within each aggregate. Another interesting property of F-SAM is that it can eliminate bias against short TCP flows in aggregate traffic consisting of both long and short TCP flows. Short lived flows do not get enough time to ramp up their sending rate and RTT estimation, and any packet loss in short flows can be lethal. In [1] and [10], the authors look at ways to remove this bias against short TCP flows in the marking and queueing domains. However these papers use per-flow mechanisms to reduce this bias, while F-SAM is a completely stateless marker and reduces this bias by the very nature of its probabilistic marking based on average rate of each flow. To provide a theoretical explanation of how F-SAM preferentially treats short flows, one needs to keep in mind that in the CSFQ algorithm the dropping probability is inversely proportional to the flow rate specified in the packet header. Now, if we use an EWMA based rate estimation algorithm, then for short flows, the rate estimation calculation is lower than the actual sending rate of the flow, since the initial rate of the short flow is small and the EWMA does not have time to ramp up before the flow ends. Since packets belonging to short flows arriving at the F-SAM marker have a lower rate in their header than their actual sending rate, the short flows gets a slightly bigger proportion of IN packets than what they are actually entitled to in a fair queueing scheme. This translates into better throughput for the short flows, as we shall see in the next.

Fair Stateless Aggregate Traffic Marking

6

217

Experimental Results

d1

s1

s2 d2 s3

e1

c (core)

e2

d3

s4

Fig. 1. Simulation scenario: A simple core with two edges. The si inject traffic of different classes into the network while the di s are the sinks. In this section we evaluate our two AQM-strategy based markers, PAM and F-SAM in progressively complex scenarios, using packet level network simulation, and compare their performance with aggregate markers such as the TokenBucket marker(TB) and the TSW2CM marker [8]. 6.1

Experimental Setup

We have used the ns-2 [20] network simulator. The topology used in our experiments is depicted in Figure 1. The source nodes are si and the destination or sink nodes are di . The source nodes are connected to the edge node e1 which inject traffic into the network core, c. The core is connected to the edge 2e which is further connected to the sink nodes di s. The source node s3 is used to generate background traffic in form of many TCP flows carrying bulk traffic. This background traffic is marked with a separate DSCP (e.g., 20) and is absorbed at d2 . 6.2

Validating the Marking Process

Marker TokenBucket PAM F-SAM

TCP Flow 1 IN pkts OUT pkts 67 1261 236 935 740 882

TCP Flow 2 IN pkts OUT pkts 76 1267 378 1456 748 835

Misbehaving UDP Flow IN pkts OUT pkts 2423 7215 2011 7607 1042 8254

Fig. 2. Detailed Packet Marking results with 2 TCP flows and 1 misbehaving UDP flow. For every marker, we give the number of IN and OUT packets obtained by each flow at the marker. The UDP flow has a rate of 2Mb/s.

First, we evaluate our PAM and F-SAM markers and check the amount of tokens obtained by individual flows in an aggregate. We conduct simulations for

218

A. Das, D. Dutta, and A. Helmy

Marker TokenBucket PAM F-SAM

UDP Flow 1 (1Mb/s) IN pkts OUT pkts 13 748 245 695 497 620

UDP Flow 2 (4Mb/s) IN pkts OUT pkts 805 2869 695 2883 413 2952

UDP Flow 3 (2Mb/s) IN pkts OUT pkts 387 1593 387 1593 456 1571

Fig. 3. Detailed Packet Marking Results with three UDP flows of varying bandwidths. For every marker, we give the number of IN and OUT packets obtained by each flow at the marker. a mix of UDP and TCP flows and for a mix of UDP flows of varying bandwidth. We verify that the proportion of packets marked for each flow by our marker conforms to our claims. This result helps us to estimate the end-to-end performance results. Due to lack of space, we present a very brief summary. A detailed description can be found in [4]. Consider the table in Figure 2. This experiment had 2 TCP flows multiplexed with 1 misbehaving UDP flow(2Mb/s) injected into the core. For our markers, we use a committed information rate (CIR) of 500kb/s and a burst size of 100kb and ran the simulation for 40 seconds in virtual time. We also use background TCP traffic in the topology to prevent flow synchronization. We calculate the number of IN and OUT packets marked for each flow and we find that TokenBucket markers allows the misbehaving UDP flow to squeeze the elastic TCP connections of IN tokens. However, in case of PAM, as we increase the sending rate, PAM marks IN packets proportionately i.e. the misbehaving flow gets penalized more (in terms of getting lesser IN packets than in the TB case). In other IN packets words, with PAM the ratio total packets transmitted is approximately the same for all flows. F-SAM demonstrates a fairer token allocation. We see that, in F-SAM, all the flows get approximately similar share of tokens from the marker, in spite of the large misbehaving UDP flow (which misappropriated most of the tokens from the TCP flows in the case of the TokenBucket marker). Thus we see, that the F-SAM marker effectively segregates flows from each other, and does not let misbehaving flows take away tokens from other well-behaved flows. In the next experiment, we consider only three UDP flows carrying CBR traffic, with varying transmission rates (1Mb/s, 4Mb/s, and 2Mb/s). Looking at Figure 3, one can see a similar improvement of PAM and FSAM over a TokenBucket Marker with respect to the fairness of token allocation among the three flows. With F-SAM, the number of tokens allocated to each flow is nearly the same, in spite of widely varying individual flows rates. Comparing with the token-bucket marker results , where the larger flows gather most of the tokens, we see that the F-SAM marker effectively isolates flows among the aggregates. It also evenly distributes tokens among the individual flows in a max-min fair manner, without maintaining any per-flow state. 6.3

End-to-End Performance

In this subsection, we compare the end-to-end performance of PAM and F-SAM with the TB (TokenBucket marker) and the sophisticated TSW2CM marker.

Fair Stateless Aggregate Traffic Marking

TB PAM F-SAM TSW2CM

20 Per-flow throughput (pkts/sec)

219

10

0

0

50

100 Number of bulk TCP flows

150

Fig. 4. Average throughput of flows, when only bulk TCP flows are used between s1

and d1

40 Token Bucket PAM F-SAM TSW2CM

Token Bucket PAM F-SAM TSW2CM

35

Per-flow throughput (pkts/sec)

Per-flow throughput (pkts/sec)

20

10

30 25 20 15 10 5

0

0

50

100 200 150 Number of bulk TCP flows

250

300

1em

0

0

(i) Elephants

50

100 200 150 Number of bulk TCP flows

250

300

(ii) Mice

Fig. 5. Average throughput of flows when both bulk (elephants) as well as short lived TCP flows (mice) are used between s1 and d1

In all the experiments in this subsection, the CIR was 1Mb/s, the TB burst size was 500Kb/s and there was no separate background traffic as in the previous set of experiments. All the links except the one between core and e2(the bottleneck link) had a bandwidth of 10Mb/s, while the bottleneck had 5Mb/s. The delays in all the links were 5ms. We have used TCP Reno for all TCP agents and the short lived flows were generated by the HTTP 1.0 model in ns-2 . The parameters for the RIO queues were the default ones in ns-2 . We have got better results by tweaking the RIO parameters but we have not presented those results here. First, we test our markers with long lived bulk TCP flows from s1 to d1 shown in Figure 4. We see that F-SAM is around 10% better than TB. PAM too performs better than TB (by 2-5%). It is interesting to note that both the throughputs due to PAM as well as F-SAM have a lesser standard deviation compared to TB, implying that our markers are more fair. The clear winner is F-SAM which has a much lesser standard deviation as it uses a max-min fairness criteria to distribute the token at the marker. Note that TSW2CM performs worse than TB in this scenario.

220

A. Das, D. Dutta, and A. Helmy

In the next experiment, we take several bulk TCP flows along with several short lived TCP flows (with an average arrival rate of 1 second and a payload size of 20KB) marked from the same token bucket. We illustrate the results for the long flow and short flows separately. Again, the average per-flow throughput for long flows with PAM and F-SAM yielded higher throughput and less standard deviation (Figure 5(i)). We must note that the F-SAM results have very low standard deviation (almost half of that of the TB marker), which demonstrates the fairness of F-SAM marking. Figure 5(ii) depicts the bandwidth obtained by the short flow aggregates. Again, F-SAM beats the competition by a large margin. This is important as we can use F-SAM to win the war between mice and elephants[10] without maintaining per-flow state. Note that we do not need to detect which flows are short as our max-min fairness based F-SAM ensures fair token distribution. Also note that, the bandwidth obtained by mice is almost constant unlike the other markers. Surprisingly, for the short flows, we see that TB is better than PAM.

140

Per-flow throughput (pkts/sec)

Bandwidth of misbehaving UDP flow (pkts/sec)

Token Bucket PAM F-SAM TSW2CM

40

30

20

10

0

0

50

100 Number of bulk TCP flows

(i) TCP flows

150

1em

TB PAM F-SAM TSW2CM

135 130 125 120 115 110 105 100 95 90

0

50

100 Number of bulk TCP flows

150

(ii) Misbehaving flow

Fig. 6. Bandwidth obtained by flows when bulk TCP flows are present between s1 and d1 along with a single misbehaving UDP flow with a rate of 1Mb/s between s2 and d1 In the next set of experiments, we use a single misbehaving UDP flow generating CBR traffic at 1 Mb/s apart from many bulk TCP flows. The throughput results for the TCP flows are depicted in Figures 6(i), (ii). As a result of our marking, we see that the average bandwidth received by the bulk TCP flows in F-SAM is the highest among all the markers(around 50% more than TB). PAM too is higher than TB by around 30% and is marginally better than TSW2CM. We note that the performance of TSW2CM is closer to PAM since probabilistic rate based marking in TSW2CM too helps to check misbehaving UDP flows. The better TCP performance of F-SAM is clearly due to an evenly distributed outprofile packets (packets marked OUT) compared to bursty out-profile packets in the case of the TB marker in presence of misbehaving UDP flows. This is also demonstrated by Figure 6(i) which shows the bandwidth obtained by the misbehaving UDP flow in the previous scenario. Clearly this result correlates with

Fair Stateless Aggregate Traffic Marking

1

221

1 TB PAM F-SAM TSW2CM

0.9

TB PAM F-SAM TSW2COL

0.9 0.8

Fairness Index

Fairness Index

0.7 0.8

0.7

0.6 0.5 0.4 0.3

0.6

0.2 0.1

0.5

0

100 50 Number of bulk TCP flows along with a misbehaving UDP flow

(i) AllTCP + 1UDP

150

0

0

5

10 Number of UDP flows

15

(ii) All UDP flows

Fig. 7. Comparison of the Fairness Index of markers when only (i) many TCP flows and 1 misbehaving UDP flow is present and (ii) When only UDP flows are present.

Figure 6(ii), as we see F-SAM (and PAM to a lesser extent) penalizing the UDP flow much more than a TB marker. The fairer token distribution ensures that bandwidth is taken away from the misbehaving flow and is distributed among the TCP flows. Figure 7(i) plots the fairness index for the previous experiment using the throughput of the TCP flows and the UDP flow. The fairness index [5] can be written as N ( i xi )2 (2) FI = N N × i x2i where xi is the per-flow throughput. We see a marked increase in the fairness index of F-SAM over all the other markers. PAM too exhibits a much greater fairness index than TB. The performance improvement for both F-SAM and PAM continues as the bandwidth of the misbehaving flow is increased. This demonstrates the efficacy of our marking schemes. We should note that no special tuning of queue parameters were required to get our performance improvement. In fact, we have obtained better performance when we tuned the parameters of the PAM marker as well as the RIO queues. Next, we consider a scenario with only UDP flows. This is important for multimedia transmission, where we want well behaved multimedia streams to be protected by misbehaving flows. In this experiment, we used an aggregate consisting of a varying number of UDP flows, each with a different bandwidth varying from 2Mb/s to 10Mb/s. Our results show that while TB shows a high degree of unfairness in the allocation of tokens to the different UDP flows, FSAM allocates approximately the same number of tokens to all the UDP flows irrespective of their bandwidth. PAM (and TSW2CM) try and allocate tokens in proportion to the incoming rate and also show greater fairness than TB. Figure 7(ii) plots the fairness index among UDP flows within the aggregate. The figure clearly shows that F-SAM treats flows fairly irrespective of their bandwidth and hence protects well behaved multimedia flows from large misbehaving

222

A. Das, D. Dutta, and A. Helmy

flows. We see that while TB performs the worst, PAM and TSW2CM perform better than TB and behave similar to each other. Thus, we demonstrate that F-SAM does well to isolate misbehaving flows. We conjecture that techniques such as F-SAM perform close to the upper bound for fairness in aggregate marking, without per-flow calculations. PAM is a much less sophisticated marker, has less dramatic gains but its easier to implement than F-SAM. But they are both very easily deployable since they are stateless. Their performance shows a marked improvement over simplistic TokenBucket aggregate markers, and in fact come closer to that of many sophisticated perflow marking schemes discussed in Section 2.

7

Conclusion and Future Work

In this paper, we have proposed two AQM based aggregate markers that ensure fairness among different flows in the aggregate without maintaining per-flow state. This makes our approach unique. Probabilistic Aggregate Marking (PAM) ensures fairness in a proportional fashion and allocates tokens to packets with a probability transfer function that looks similar to the transfer function of RED. We also presented a more sophisticated marker called Fair Stateless Aggregate Marker (F-SAM) which is based on fair queueing using max-min fairness criteria. The promising aspect of our work is F-SAM which can boost the performance of short lived flows and also ensure fairer bandwidth allocation between TCP and UDP flows of an aggregate, and within TCP aggregates and UDP aggregates separately. The above markers are scalable and readily deployable. Our hope is that markers like PAM and F-SAM will enable a much more fairer and scalable solution to the problem of traffic marking in the differentiated services framework. Future work in this area involves adapting other AQM techniques like CHOKe to develop fair aggregate markers. Additionally, for PAM, we plan to incorporate marking schemes based on PI controllers and try to dynamically auto-configure the RED-like parameters based on the traffic pattern. In F-SAM, we have not yet looked at the idea of weighted fair queueing while distributing tokens among the flows. Allowing for different weights to individual flows of the aggregate is the next logical step after approximating fair queueing. Finally, we want to explore how our markers can be used to implement good, fair pricing schemes in a diffserv QoS architecture.

References 1. A.Feroz, A. Rao, and S. Kalyanaraman. A tcp-friendly traffic marker for ip differentiated services. Proc. of the IEEE/IFIP Eighth International Workshop on Quality of Service - IWQoS, 2000. 212, 213, 216 2. S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, and W. Weiss. An architecture for differentiated services. RFC 2475, 1998. 211

Fair Stateless Aggregate Traffic Marking

223

3. Mark Claypool and Jae Chung. Dynamic-cbt and chips - router support for improved multimedia performance on the internet. ACM Multimedia, 2001. 213 4. A. Das, D. Dutta, and A. Helmy. Fair stateless aggregate marking techniques for differentiated services networks, university of southern california technical report usc-cs-tr-02-752, 2002. 216, 218 5. Bruce S. Davie and Larry L. Peterson. Computer Networks: A Systems Approach. Morgan Kauffmann, second edition, 2000. 221 6. A. Demers, S. Keshav, and S.J. Shenker. Analysis and simulation of a fair queueing algorithm. Sigcomm, 1989. 214 7. B. Braden et al. Recommendations on queue management and congestion avoidance in the internet. RFC 2309, 1998. 212 8. W. Fang, N. Seddigh, and B. Nandy. A time sliding window three colour marker (tswtcm), 2000. 212, 217 9. S. Floyd and V. Jacobson. Random early detection gateways for congestion avoidance. IEEE/ACM Transactions on Networking, V.1 N.4, 1993. 212, 213, 214, 214 10. L. Guo and I. Matta. The war between mice and elephants. ICNP, 2001. 216, 220 11. J. Heinanen and R. Guerin. A single rate three color marker. RFC 2697, 1999. 212 12. J. Heinanen and R. Guerin. A two rate three color marker. RFC 2698, 1999. 212 13. J. Heinehan, T. Finner, F. Baker, W. Weiss, and J. Wroclawski. Assured forwarding phb group. RFC 2597, 1999. 211 14. Lloyd Wood Ilias Andrikopoulos and George Pavlou. A fair traffic conditioner for the assured service in a differentiated services internet. Proceedings of ICC 2000, vol. 2 pp. 806-810, 2000. 213 15. Dong Lin and Robert Morris. Dynamics of random early detection. SIGCOMM ’97, pages 127–137, september 1997. 214 16. K. Nichols, V. Jacobson, and L. Zhang. A twobit differentiated services architecture for the internet. RFC 2638, 1999. 211 17. R. Pan, B. Prabhakar, and K. Psounis. A stateless active queue management scheme for approximating fair bandwidth allocation. IEEE INFOCOM 2000, 2000. 212, 214 18. Ion Stoica, Scott Shenker, and Hui Zhang. Core-stateless fair queueing: Achieving approximately fair bandwidth allocations in high speed networks. Sigcomm, 1998. 212, 213, 214, 215, 215, 216, 216 19. H. Su and Mohammed Atiquzzaman. Itswtcm: A new aggregate marker to improve fairness in difserv. Globecomm, 2001. 213 20. UCB/LBNL/VINT. The NS2 network simulator, available at http://www.isi.edu/nsnam/ns/. 217 21. Ikjun Yeom and A. L. Narasimha Reddy. Adaptive marking for aggregated flows. Globecomm, 2001. 212, 213 22. Lixia Zhang, Steve Deering, Deborah Estrin, Scott Shenker, and Daniel Zappala. Rsvp: A new resource reservation protocol. IEEE Network Magazine, September 1993. 211

A Dynamic Marking Scheme of Assured Service for Alleviating Unfairness among Service Level Agreements Seung-Joon Seok1, Seok-Min Hong1, Sung-Hyuck Lee2, and Chul-Hee Kang1 1 Department

of Electronics Engineering, Korea University 1, 5-ga, Anam-dong, Sungbuk-gu, Seoul 136-701 Korea {ssj, mickey, chkang}@widecomm.korea.ac.kr http://widecomm.korea.ac.kr/~ssj 2 Network Protocol T.G., i-Networking Lab., Samsung Advanced Institute of Technology San 14-1, Nongseo-ri, Kiheung-eup, Yongin-shi, Kyungki-do, Korea [email protected]

Abstract. Assured Service, which is a service model of the Internet Differentiated Services (DiffServ) architecture, is not currently well implemented on the Internet, mainly because TCP employs an AIMD (Additive Increase and Multiplicative Decrease) mechanism to control congestion. Many studies have shown that the current Assured Service model does not assure the target rate of high profile flows in the presence of numerous low profile flows, and does not equably distribute the reservation rate of an SLA among multiple flows included in the SLA. In this paper, the former problem is referred to as inter-SLA unfairness and the latter problem is referred to as intra-SLA unfairness. We propose a marking rule, called RAM (Rate Adaptive Marking) that simultaneously diminishes both of these unfairness problems. The RAM method marks sending packets, at a source node or at an edge router, in inverse proportion to throughput gain and in proportion to the reservation rate and throughput dynamics. Three experiments with the ns-2 simulator are performed to evaluate the RAM scheme. The simulation results show that the RAM scheme may significantly reduce inter-SLA unfairness and the intra-SLA unfairness.

1 Introduction An Assured Service (AS), as first defined in [1], is an example of an end-to-end service that can be built from the proposed differentiated services enhancements to IP using a single PHB of Assured Forwarding PHB. The basic concept of the AS is that packets are forwarded with high probability as long as the traffic from the customer site does not exceed the subscribed rate (traffic profile). Thus, it is necessary for the Assured Service that the interior router allocates some bandwidth and a buffer for each AF class, and that a user or user group of an Assured Service establishes a contract with the provider, which defines the profile for the service expected to be obtained; SLA (Service Level Agreement). This contract can be made only for a TCP flow, aggregated TCP flows or a combination of multi-protocol flows. The building blocks of the AS include a traffic marker at the edge router of a domain and a differThe original version of this chapter was revised: The copyright line was incorrect. This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 224-236, 2002. © IFIP International Federation for Information Processing 2002

A Dynamic Marking Scheme of Assured Service

225

entiated dropping algorithm in the network interior router. The traffic marker ranks packets as IN or OUT depending upon the SLA. An example of a differentiated dropping algorithm is RIO (RED with IN and OUT). The RIO algorithm uses the same RED algorithm for the IN packets and OUT packets, albeit with a different set of parameters for each. In particular, the OUT packets are preferentially dropped, upon evidence of congestion at the bottleneck, before the IN packets. In previous studies, it was reported that there are two types of unfairness in the case of an Assured Service. Firstly, let us suppose that TCP flows with an Assured Service path through the same bottleneck link. In particular, let us assume that the TCP flows have different traffic profile rates. Under these conditions, the flows expect to share the excess bandwidth of the bottleneck link so as to meet their target rates. The target rate is the sum of the traffic profile rate plus an equal share of the excess bandwidth of the bottleneck link for all flows. This is called the equal sharing target rate [3]. However, several studies [4][5][6] have shown that TCP flows with a high profile rate and/or long round trip time (RTT) barely meet their target rate in the presence of numerous flows with low profile rate. The reason for this problem is that current TCP additively increases its window size and multiplicatively decreases it when any packet is lost. This causes flows with a high profile rate to take a longer time to reach to their target rates following packet loss, or not to reach to. For convenience, this problem is herein referred to as ìinter-SLA unfairnessî. Unfairness is defined as excess bandwidth of a bottleneck link not being distributed equably among flows that go through that link. Next, the SLA may cover a set of flows. In this case, however, there exists unfair sharing of the reserved bandwidth or of the profile rate among the aggregated flows [7][8][9][10]. This unfairness between the aggregated flows is a serious problem in the Assured Service model. The unfairness can be caused by differences in RTTs, in link capacities, or in congestion levels experienced by flows within the network, though the total throughput of the aggregation still reaches the reservation rate. This type of unfairness is herein referred to as ìintra-SLA unfairnessî. To balance or fix the unfairness, a sophisticated marking mechanism is required to distribute the profile rate fairly to the aggregated flows at the edge routers [7][8][9][10]. Most studies of Assured Service focus on TCP protocol extensions [3][11], shaping [12][13] and dropping policies [3][14]. However, few approaches propose to change the current general marking rule which is that the total IN marking rate is equal to the reservation rate. In this paper, we present a new marking rule to resolve the two unfairness problems. The rest of the paper is organized as follows. Section 2 describes TCP dynamics in a differentiated services network, previous approaches to changing the general marking rule in order to alleviate inter-SLA unfairness and the axiom of the approach set out in this paper, called the Rate Adaptive Marking (RAM) scheme, in order to reduce the dynamics. Section 3 details the architectures and algorithms of the RAM. Section 4 studies the performance of RAM using an ns-2 network simulator and compares it with that of normal marking schemes and section 5 discusses some unresolved issues and concludes this paper.

226

S.-J. Seok et al.

(a) general standard marking scheme

(b) Fengís soft bandwidth guarantee scheme

(c) Namís proportional marking scheme

(d) proposed RAM scheme

Figure 1. Marking methods for two flows in the cases when a high profile flows and a low profile flows competes with each other at a bottleneck link, marking methods.

2 Two Previous Works and Rate Adaptive Marching Scheme In the Assured Service architecture, it is generally assumed that sending packets are marked with IN as long as they do not exceed the reservation rate (traffic profile rate), otherwise marked OUT, since the marking rate is the same as the profile rate of the SLA. Figure 1(a) shows this general marking scheme for two flows having different reservation rates. In this figure, the shadowed region indicates the quantity of OUT marked packets. So far, there are few solutions that approach the problem from the point of view of changing this general marking rule. In this section, we describe the two approaches which do so, and then propose a new approach, called rate adaptive marking. 2.1 Previous Approaches Wu Cheng Feng [15] proposed a ìsoft bandwidth guaranteeî scheme that determines the marking rate adaptively according to a flowís throughput. In this scheme, by de-

A Dynamic Marking Scheme of Assured Service

227

fault, all packets are generated as low priority packets. If however, the observed throughput falls below the minimum requested rate, the marking mechanism starts prioritizing packets until the desired rate is reached. He considered that this scheme was carried out at the edge node or the source node. The first was referred to as source transparent marking and the second was referred to as source integrated marking. In particular, the second scheme considered the integration of the marking mechanism and the TCP control mechanism. Figure 4(b) shows marking method of Fengís scheme. Fengís scheme considers the reservation rate to be allocated to a flow as a target rate. Thus, it can not provide flows with equal share of excess bandwidth, in the case where network is lightly loaded. Nam et al. [16] use a proportional marking method in which low profile flows have more OUT marked packets and, therefore, more dropped packets. In the proportional marking scheme, when a packet arrives, a meter calculates the average rate. If this average rate exceeds the flowís target rate, the proportional marking method marks the code point of the packet as OUT in proportion to the difference between the average rate and the target rate. Figure 4(c) shows the marking method of Namís scheme. As figure 4(c) shows, the high profile flow has a marking rate which is lower than its reservation rate, following a situation in which temporary throughput is over the reservation rate. This can not help the fairness, though decreasing the marking rate of the low profile may help. In this paper, another marking scheme, called Rate Adaptive Marking (RAM), is proposed. The two previous schemes considered only inter-SLA unfairness, while the RAM scheme can alleviate both inter-SLA unfairness and intra-SLA unfairness. The details of this scheme are described in the next two sections. 2.2 The Axiom of Rate Adaptive Marking What is the Rate Adaptive Marking (RAM)? This subsection describes the conceptual and operational mechanisms of this scheme. A RAM marker determines a flowís marking rates periodically, depending on information about throughput rates collected during a previous period, and marks sending packets according to this rate during the next period. The basic principle of RAM strategy is marking packets as IN in inverse proportion to the throughput gain and in proportion to the reservation rate (equation1, 2). The gain denotes how much the throughput of a flow is over its reservation rate. Also, RAM makes a flowís IN marking rate to be proportional to the dynamics in that flowís throughput (equation 3). Unless otherwise stated, a simple marking rate denotes the IN marking rate. 1 ∝ marking _ rate throughput _ gain

(1)

reservatio n _ rate ∝ marking _ rate

(2)

throughput _ dynamics ∝ marking _ rate

(3)

228

S.-J. Seok et al.

Let us assume that two flows with the same reservation rate have different throughputs. In the case of RAM, the flow that gets the higher throughput is given a lower marking rate than the flow with lower throughput. This marking strategy causes that one flow to experience packet loss earlier than the other, when next congestion arises. Therefore, the gap between the two throughputs is reduced. Whereas, in the normal marking strategy, the two flows are given identical marking rates, because the marking rate of a flow is generally the same as its reservation rate. Next, let us assume another case where there are two flows with different reservation rates: a high profile flow and a low profile flow. The difference in reservation rate is the most important reason leading to inter-SLA unfairness. The high profile flow takes a longer time to reach the temporal throughput just before a packet loss, because the congestion window size is reduced more in the high profile flow than in the low profile flow. In other words, the high profile flow has larger throughput dynamics, between the maximum temporal throughput and the average throughput (or minimum temporal throughput), than the low profile flow. Thus, in this case, RAM tries to render the marking rate proportional to the throughput dynamics. The RAM marker maintains state information of all flows that pass through it, and periodically determines the marking rate for each flow according to the following RAM rule (equation 4). The RAM marker monitors two throughput levels for all flows; temporal throughput and long term average throughput. The temporal throughput is an average throughput for a short term interval such as an average RTT, while the long-term average throughput is based on a relative long period of time such as the congestion interval. The following equation is the RAMís rule for determining the marking rate. marking _ rate =

(Th max − Thavg ) × RsvRate (Th max − RsvRate )

(4)

where Thmax is the maximum temporal throughput during the immediately preceding congestion interval and Thavg is the long term average throughput of the flow during the congestion interval. This equation connotes the basic principles of the RAM scheme. In the right-hand side of the equation, the ìThmax-Thavgî term denotes the flowís throughput dynamics and ì(Thmax-RsvRate)/(Thmax-Thavg)î the flowís throughput gain. Finally ìRsvRateî is the flowís reservation rate. If a flow achieves an average throughput that is equal to its reservation rate (RsvRate) in the RAM scheme, the marking rate becomes the reservation rate. Also, if the average throughput is over the reservation rate, the marking rate will be lower than the reservation rate and proportional to the gap between the average throughput and the reservation rate. In contrast, if the average throughput is below the reservation rate, the marking rate will be higher than the reservation rate and also inversely proportional to the gap. However, it can happen that the result of equation 4 is negative or has an infinite value, when Thmax is below or equal to the RsvRate. This case denotes that the throughput does not reach its reservation rate as well as its target rate. The RAM scheme has an upper bound and a lower bound for the marking rate, in order to protect the marking rate and the throughput from experiencing heavy fluctuations. The upper bound is considered to be several times the reservation rate (ìα×RsvRateî) and the lower bound to be a half or less of the reservation rate





$'\QDPLF0DUNLQJ6FKHPHRI$VVXUHG6HUYLFH



³ȕ×5VY5DWH´ 6RWKHPDUNLQJUDWHLVVHWWRWKHXSSHUERXQGLIWKHUHVXOWRIHTXDWLRQ LVRYHUWKHXSSHUERXQGRUQHJDWLYHDQGWKHORZHUERXQGLI WKHUHVXOWLVSRVLWLYH DQGXQGHUWKHORZHUERXQG7KXVWKHHTXDWLRQFDQEHUHZULWWHQDVIROORZV § § § 7K PD[ − 7KDYJ · · · PDUNLQJ B UDWH = PLQ ¨¨ α  PD[¨¨ β  ¨ ¸ ¸¸ ¸¸ × 5VY5DWH  7K 5VY5DWH − PD[ ¹¹¹ © © ©

 

7KLV 5$0 VWUDWHJ\ VXSSRUWV YDULRXV QHWZRUN HQYLURQPHQWV DQG LQ SDUWLFXODU YDULRXVQHWZRUNFRQJHVWLRQOHYHOV,IWKHQHWZRUNLVRYHUVXEVFULEHGRUWKHIORZVDUH XQGHUSURYLVLRQHG WKH DYHUDJH WKURXJKSXWV RI WKH IORZV RU PD[LPXP WHPSRUDO WKURXJKSXWV 7KPD[ PD\EHEHORZWKHLUUHVHUYDWLRQUDWHVDQGVRWKHPDUNLQJUDWHVRI WKHIORZVDUHPXFKKLJKHUWKDQWKHFRUUHVSRQGLQJUHVHUYDWLRQUDWHV,QWKLVFDVHDO PRVW DOO SDFNHWV DUH PDUNHG DV ,1 KLJK SUHFHGHQFH  DQG SDFNHW ORVVHV DUH SURSRU WLRQDOWRWKHIORZ¶VPDUNLQJUDWH$VDUHVXOWWKHIORZ¶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¶VSDFNHWORVVHV7KXVWKHHGJHPDUNHUKDV WZRWLPHLQWHUYDOVWRLPSOHPHQWIRUWKH5$0VFKHPH(YHU\DYHUDJHURXQGWULSWLPH WKHHGJHPDUNHUFDOFXODWHVWKHWHPSRUDOWKURXJKSXWDVWKHDYHUDJHVHQGLQJUDWHIRUWKH

230

S.-J. Seok et al.

round trip time and compares this with previous ones, in order to determine the maximum value. Also, every average congestion interval, the average sending rate for the interval is calculated. Other behaviors of this marking scheme are the same as that of the source marking RAM scheme. It is more general that an SLA includes multiple flows and a single traffic profile of aggregated flows. In this case, a marker must be located at the edge router to support multiple flows from different sources. The general marking scheme does not maintain each flow state and marks packets as IN, in proportion to the sending rates of the flows. The RAM scheme should also be carried out at the edge router. The RAM marker generates a RAM instance corresponding to each flow. This RAM instance determines the marking rate depending on RAM strategy (equation 5) and marks packets according to this marking rate. The Reservation rate of each flow is considered as the traffic profile rate for aggregated flows / the number of flows. The

Figure 2. Experimental topology used to study inter-SLA unfairness Table 1. Simulation parameters for evaluation of interSLA unfairness Parameters S-I link bandwidth S-I link delay I-C link bandwidth I-C link delay C-E link bandwidth C-E link delay E-D link bandwidth E-D link delay Router buffer size Packet size RIO parameter 1 (IN) RIO parameter 2 (OUT) Simulation time

Value 1 Mbps 1 ms 10 Mbps 10 ms 5,6.3,7,8, and 9 Mbps 100 ms 1 Mbps 1 ms 100 packets 1000 byte 50/80/0.02 10/40/0.5 200 s

A Dynamic Marking Scheme of Assured Service

231

marking algorithm for each flow is the same as the edge marking scheme of single flow. In particular, this scheme also uses the sending rate of each flow as information to determine the marking rate of the flow.

3 Performance Study In order to confirm the performance effect of the RAM scheme as against previous approaches, two experiments have been performed, using ns-2 [20], on several simulation models. The first experiment is the evaluation of inter-SLA unfairness through the comparison of the RAM with Fengís soft bandwidth guarantee, Namís proportional marking and the general marking scheme and the second experiment tests the intra-SLA unfairness using the RAM scheme, the general marking, and equally sharing scheme. 3.1 Evaluation of Inter-SLA Unfairness We first examine the effectiveness of the RAM scheme to alleviate the inter-SLA unfairness when there are only TCP flows. Both RAM scheme types, source marking RAM and edge marking RAM, are considered in this experiment. Figure 2 depicts the

Figure 3. Average Throughputs of TCP flows using RAM implemented in TCP source and using normal marking solid line : target rate (Target) filled item : average throughput using general marking (Gen.) empty item : average throughput using source marking RAM (RAM)

232

S.-J. Seok et al.

testbed topology for this experiment. In this topology, there are 60 TCP flows and each edge router is connected with 10 TCP sources. The profile rates for each TCP flow of s1-x, s2-x, s3-x, s4-x, s5-x and s6-x are 0.01Mbps, 0.02Mbps, 0.04Mbps, 0.08Mbps, 0.16Mbps and 0.32Mbps, respectively. All flows are routed through a common bottleneck link C1-E1 to model the competition for the excess bandwidth. In this way, the total reservation rate of the bottleneck link is 6.3Mbps. Also, this experiment considers several network environments in which different bottleneck link bandwidths, x = 5, 6.3, 7, 8, and 9Mbps, are used. Other parameters of the testbed are represented in table 1.

Figure 4. Average Throughputs of TCP flows using RAM implemented in TCP source and in edge router solid line : RAM scheme in source (S_RAM) item : RAM scheme in edge router (E_RAM)

Figure 5. Comparison of effectiveness of marking schemes in the case when bottleneck link bandwidth is 9Mbps.

A Dynamic Marking Scheme of Assured Service

233

Figure 3 shows the average throughputs of the 10 TCP flows with the same reservation rate when the source marking RAM rule is applied and when the general marking rule is applied. These throughputs are compared with the ideal target rates. It is impossible for all average throughputs to be at the ideal target rates. It is required, however, for all average throughputs to be fairly near to their target rates. When the general marking rule is used, low profile (0.01, 0.02, 0.04, or 0.08 Mbps) flows achieve throughputs that are near or over their target rates, regardless of network load. Also, these throughputs are higher than those obtained when the source marking RAM rule is used. When the general marking is applied, however, the throughputs of the high profile (0.08, 0.16, or 0.32 Mbps) flows are less than those obtained when the source marking RAM rule is used, as well as being much less than the target rates. This problem can be ignored in the over-subscribed case, but is made worse as the network becomes under-loaded. However the RAM scheme can alleviate this problem. In figure 4, the source marking RAM rule is compared with the edge marking RAM rule. From this figure, we can see that the edge marking RAM technique can remedy the above problem as well as the source marking RAM technique does, because the result of the simulation using the edge marking RAM rule is similar to that of the source marking RAM rule, regardless of network load. Next, we compare the source marking RAM technique with Namís scheme, Fengís scheme and the general marking scheme, when the bottleneck link is well-provisioned (9Mbps). Figure 5 shows that Namís scheme gives the 40, 80, and 160Kbps profile flows the maximum throughput. For the 320Kbps profile flow, however, the RAM scheme is optimal. Because inter-SLA unfairness is mainly impacted by the degradation of the 320 Kbps flows (figure 3), improving the throughputs of these flows is the key factor for resolving this problem.

Figure 6. Experimental topology used to study the intra-SLA unfairness

234

S.-J. Seok et al. Table 2. Simulation parameters for evaluation of intra-SLA unfairness Parameters Value S-I or E-I link bandwidth 1 Mbps S-I or E-I link delay 1 ms I-C link bandwidth 10 Mbps I-C link delay 10 ms C-E link bandwidth 500/300 Kbps (r1, r2, r3, r4, r5) C-E link delay 100 ms (d1, d2, d3, d4, d5) E-D link bandwidth 10 Mbps E-D link delay 1 ms Router buffer size 100 packets Packet size 1000 byte RIO parameter 1 (IN) 50/80/0.02 RIO parameter 2 (OUT) 10/40/0.5 Simulation time 200 s

3.2 Evaluation of Intra-SLA Unfairness In this experiment, we consider the case in which an SLA includes multiple flows. Figure 6 depicts the experimental topology used to test the intra-SLA unfairness. In this topology, domain 1 and domain 2 establish an SLA, in which 500Kbps is allocated as the traffic profile of the aggregated flows from sources (s1, s2, s3, s4 and s5). Thus, these flows should be remarked at the E1 according to the SLA. The other flows are used for background traffic and assumed to be coming into domain 2 from different domains, except for domain1. The background sources (s6, s7, s8, s9, and s10) have different marking rates (10, 100, 200, 300, and 400 Kbps respectively). In order that each flow included in the SLA experiences different network status, flows1, 2, 3, 4 and 5 are routed through different routes with different congestion levels within domain 2. The congestion levels are controlled by the background flows. Other simulation parameters are represented in table 2. In the ideal case, each flow included in an SLA has a throughput and Jainís fairness index becomes one. As unfairness among the flows increase, the fairness index value decreases. In this experiment, we carried out two simulations. In the first simulation, CBR streams are considered as background traffic, and in the second simulation, TCP streams are considered. We monitored the average throughputs of TCP flows included in the SLA, in the case when general marking, RAM-edge marking, or equal sharing marking (Sharing) is used. The general marking scheme marks incoming packets in proportion to the incoming rate of each flow. In equal sharing marking, however, the SLAís reservation rate is distributed equally among aggregated TCP flows. Figure 7 shows the simulation results. In first simulation result (figure 7(a)), the fairness index value of the general, RAM, and Sharing scheme are about 0.88, 0.95, and 0.93, respectively. In the second simulation result (figure 7(b)), these values are 0.93, 0.98, and 0.97, respectively. From these results, we confirmed that the RAM scheme can reduce intra-SLA unfairness as well as inter-SLA unfairness.

A Dynamic Marking Scheme of Assured Service

235

(a) In the environment in which background traffic is CBR stream, all core-egress link bandwidth is 0.5 Mbps.

(b) In the environment in which background traffic is TCP stream, all core-egress link bandwidth is 0.35 Mbps Figure 7. Average Throughputs of TCP flows included in a SLA

4 Conclusion In this paper, we described two serious problems, inter-SLA unfairness and intraSLA unfairness, which arise when providing an Assured Service and we discussed previous reports on this topic. We proposed an approach which consists of a modified marking rule, called RAM (Rate Adaptive Marking), for the Assured Service, to alleviate both unfairness problems. The RAM scheme determines a flowís marking rate periodically, in inverse proportion to throughput gain, and in proportion to throughput dynamics and to the reservation rate during a previous time period. We described two scenarios to implement the RAM scheme, source marking RAM and edge marking RAM. The source marking RAM scheme cooperates with TCP congestion control to calculate throughputs accurately. The edge marking RAM scheme operates at an ingress edge router and exploits the sending rate of each flow, instead of the throughput. The edge marking RAM scheme can support the marking of aggregated flows, whereas the source marking RAM scheme can not. The effectiveness of RAM was evaluated through computer simulations. The simulation results showed that the RAM technique can alleviate both inter-SLA unfairness and intra-SLA unfairness simultaneously.

236

S.-J. Seok et al.

References 1. Heinanen, J., Baker, F., Weiss, W., and Wroclawski, J.: Assured Forwarding PHB Group. RFC2597, June (1999) 2. Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z., Weiss, W.: An architecture for differentiated services. RFC 2598, June(1999) 3. Yeom, I., Reddy, A. L. N.: Realizing Throughput Guarantees in a Differentiated Service Network. In Proc. ICMCSí99, June (1999) 372-376 4. Rezende, J. F.: Assured Service Evaluation. In Proc. GLOBECOMí99, June (1999) 100-104 5. Seddigh, N., Nandy, B., Pieda, P.: Bandwidth Assurance Issues for TCP flows in a Differentiated Services Network. In Proc. GLOBECOMí99, June (1999) 1792-1798. 6. Ibanez, J. Nichols, K.: Preliminary Simulation Evaluation of an Assured Service. Internet Draft, Aug.(1998) 7. Yeom, I., Reddy, A. L. N.: Impact of marking strategy on aggregated flows in a differentiated services network. In Proc. IWQoS 1999 , May(1999) 8. Kim, H.: A Fair Marker. Internet Draft, April(1999) 9. Andrikopoulos, L., Pavlou, G.: A fair traffic conditioner for the assured service in a differentiated service Internet. In Proc. IEEE ICC 2000, June(2000) 10. Azeem, F., Rao, A., Kalyanaraman, S.: TCP-friendly traffic marker for IP differentiated services, In Proc. IWQoS'2000, June(2000) 35-48 11. Fang, W., Perterson, L.: TCP mechanisms for a diff-serv architecture, Tec. Rep 12. Cnodder, S. D., Elloumi, O., Pauwels, K.,: Rate adaptive shaping for the efficient transport of data traffic in diffserv networks. Computer Networks, Vol. 35(2001) 263-285 13. Li, N., Borrego, M., Li, S.: A rate regulating traffic conditioner for supporting TCP over Diffserv. Computer Commnications, Vol. 23(2000) 1349-1362 14. Park, W. H., Bahk, S., Kim, H.: A modified RIO algorithm that alleviates the bandwidth skew problem in Internet Differentiated Service. In Proc. ICCí00 (2000) 1599-1603 15. Feng, W., Kandlur, D., Saha, D., Shin, K.: Adaptive Packet Marking for Providing Differentiated Services in the Internet. In Proc. ICNPí98, Austin, TX, Oct. (1998) 108-117 16. Nam, D-H., Choi, Y-S. Kim, B-C., and Cho, Y-Z.: A traffic conditioning and buffer management scheme for fairness in differentiated services. in proc. ICATM 2001, April (2001) 17. Baines, M.: Using TCP Model To Understand Bandwidth Assurance in a Differentiated Services Network, In proc. IEEE Infocomm 2001 (2001)1800-1805 18. Padyhe, J., Firoiu, V., Townsley, D., Kurose, J.: Modeling TCP Throughput: A Simple Model and its Empirical Validation. CMPSCI Technical Report TR 98-008, University of Massachussetts, MA (1999) 19. Lee, S-H., Seok, S-J., Lee, S-J., Kang, C-H.: A Study of TCP and UDP flows in a Differentiated Services Network using Two Markers System. In proc. MMNS 2001, Oct. (2001) 198-203 20. Network Simulator ñ NS (version 2), http://www-mash.ca.berkely.edu/ns/

Minimizing Transmission Costs through Adaptive Marking in Differentiated Services Networks Chen-Khong Tham and Yong Liu Department of Electrical and Computer Engineering National University of Singapore, Singapore 119260 {eletck,engp1130}@nus.edu.sg

Abstract. The issue of resource management in multi-domain Differentiated Services (DiffServ) networks has attracted a lot of attention from researchers who have proposed various provisioning, adaptive marking and admission control schemes. In this paper, we propose a Reinforcement Learning-based Adaptive Marking (RLAM) approach for providing end-to-end delay and throughput assurances, while minimizing packet transmission costs since ‘expensive’ Per Hop Behaviors (PHBs) like Expedited Forwarding (EF) are used only when necessary. The proposed scheme tries to satisfy per flow end-to-end QoS through control actions which act on flow aggregates in the core of the network. Using an ns2 simulation of a multi-domain DiffServ network with multimedia traffic, the RLAM scheme is shown to be effective in significantly lowering packet transmission costs without sacrificing end-to-end QoS when compared to static and random marking schemes. Keywords: Multimedia network traffic engineering and optimization; QoS management; End-to-end IP multimedia network and service management

1

Introduction

Users of networked applications may be willing to pay a premium to enjoy network service that is better than the best effort service found in most networks and the Internet today. However, apart from specialized applications requiring a guaranteed service [1], such as a real-time control application, most users and their generally adaptive applications usually only have loose requirements such as “low delay” or “high throughput”, perhaps with specified tolerable upper and lower limits. The Differentiated Services (DiffServ or DS) framework [2] introduced the concept of Per Hop Behaviors (PHBs) such as Expedited Forwarding (EF) [3] and Assured Forwarding (AF) [4] at different routers in DS domains with the aim of providing quality of service (QoS) assurances for different kinds of traffic. DiffServ is itself a simplification of the per-flow-based Integrated Services (IntServ) model and deals with flow aggregates instead of individual flows in the core of the DS domain and in intermediate DS domains between source and The original version of this chapter was revised: The copyright line was incorrect.This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 237–249, 2002. c IFIP International Federation for Information Processing 2002

238

C.-K. Tham and Y. Liu

destination nodes. The question arises as to what is the appropriate PHB to use at each DS domain in order to achieve a certain level of end-to-end QoS1 . Since PHBs are applied on packets based on their DiffServ Code Point (DSCP) value in the DS field of the IP packet header, the issue then becomes how to select the DSCP marking in packets belonging to flows with specific end-to-end QoS requirements. The common widely-accepted way of doing this is to mark packets from flows with stringent QoS requirements with the DSCP value corresponding to the EF PHB, packets from flows with less stringent QoS requirements with a DSCP value corresponding to a class of the AF PHB, and finally packets from flows with no specific QoS requirement with the DSCP value corresponding to the BE (best effort) PHB. To achieve some level of QoS assurance, different DS domains have Service Level Agreements (SLAs) with their neighboring DS domains which specify performance parameters or limits for traffic carried between the domains - however, these are usually in terms of worst case values which may be significantly different from what is encountered during actual use. Furthermore, the actual QoS achieved between an ingress-egress router pair in different DS domains, for a particular PHB or Per Domain Behavior (PDB) selected based on the DSCP value, may be different. In this paper, we propose a Reinforcement Learning-based Adaptive Marking (RLAM) scheme to mark packets of particular types of flows in different DS domains with the appropriate DSCP values to select specific PHBs so as to achieve the desired level of end-to-end QoS in a cost effective manner. The proposed method observes the effect on end-to-end QoS when different PHBs are selected in different DS domains in order to arrive at a PHB selection strategy at each domain for different types of flows, given the condition of the network traffic at that time. The RLAM scheme inter-operates with the underlying low-level QoS mechanisms such as scheduling and admission control, so long as they operate in a consistent and predictable manner. However, there is an implicit assumption that the desired end-to-end QoS can actually be achieved by using different PHBs in each DS domain. This assumption is not true when, for example, too much traffic has been allowed into the network which results in severe congestion and high delays and losses for all packets regardless of the selected PHB. Hence, buffer management and admission control mechanisms should also be deployed. The organization of this paper is as follows. In the next section, we survey some existing work in adaptive marking. In Section 3, we describe the theory behind the feedback- and experience-based learning control method known as reinforcement learning (RL) or neuro-dynamic programming (NDP). In Section 4, we describe the design and implementation considerations of the proposed Reinforcement Learning-based Adaptive Marking (RLAM) scheme. This is followed by the description of an ns2 implemention of RLAM in Section 5 and the presentation of simulation results in Section 6. Finally, we conclude in Section 7.

1

We simply use the term “QoS” to refer to the most common QoS parameters such as delay, jitter, throughput and loss.

Adaptive Marking in Differentiated Services Networks

2

239

Adaptive Marking in DiffServ Networks

There are two types of packet marking which take place concurrently in the DiffServ architecture: (1) marking of in- and out-of-profile packets within the traffic conditioner found in ingress or egress edge routers, and (2) marking packets with DSCP values in order to achieve the desired packet forwarding behavior at the routers in a DS domain. In the first type of marking, a meter within the traffic conditioner measures packets belonging to a flow and compares them against a traffic profile. If the packets are found to be out-of-profile, they will be marked as such for subsequent handling by the shaper which delays the packets, or the dropper which discards the packets. Alternatively, these packets can also be remarked with another DSCP value corresponding to a lower PHB. In recent literature, an interesting example of this type of marking can be found in [5] in which a three-colour marking scheme in a Random Early Demotion and Promotion (REDP) marker allows EF or AF packets which have been demoted when the agreed bandwidth between certain domains have been exceeded, to be promoted again to their original marking so that they will be served ahead of BE packets in domains which have available bandwidth. The second type of marking is the common mode of operation in DiffServ networks, in which either the source, a leaf router in the source domain, or the first ingress edge router encountered by the packet, provides the DSCP marking which usually remains unchanged all the way to the destination. In this paper, we focus on an adaptive form of this second type of marking which will be done even for in-profile packets. An application of dynamic marking is described in [6] where a Packet Marking Engine (PME) marks with high priority packets from important TCP flows that will otherwise fall below their required throughput due to competition with other flows.

3

Reinforcement Learning

Reinforcement learning (RL) [7] (also known as neuro-dynamic programming (NDP) [8]) is a form of machine learning in which the learning agent has to formulate a policy which determines the appropriate action to take in each state in order to maximize the expected cumulative reward over time. An effective way to achieve reinforcement learning is to use the Q-Learning algorithm [9] in which the value of state-action pairs Q(x, a) are maintained and updated over time in the manner shown in Equation 1:   Qt (x, a) + ηt [rt + γVt (yt ) − Qt (x, a)] (1) Qt+1 (x, a) = if x = xt and a = at ,   Qt (x, a) otherwise. where yt is the next state when action at is taken in state xt , Vt (yt ) = maxl∈A(yt ) Qt (yt , l), A(yt ) is the set of available actions in state yt and rt is the immediate

240

C.-K. Tham and Y. Liu

reinforcement2 that evaluates the last action and state transition. The γ term discounts the Q-value from the next state to give more weight to states which are near in time since they are more responsible for the observed outcome, while ηt is a learning rate parameter that affects the convergence rate of the Q-values in the face of stochastic state transitions and rewards. A variation of Equation 1 is to use the Q-value associated with the actual action lt selected in state yt rather than the maximum Q-value across all actions in state yt . In this case, the Q-value update equation becomes:   Qt (x, a) + ηt [rt + γQt (yt , lt ) − Qt (x, a)] (2) Qt+1 (x, a) = if x = xt and a = at ,   otherwise. Qt (x, a) The action a in each state xt is selected according to the Boltzmann probability distribution: eβQt (xt ,a) (3) P (a|xt ) =  βQt (xt ,l) l∈A(xt ) e where A(xt ) is the set of available actions at state xt and β is a parameter which determines the probability of selecting non-greedy actions. In the area of QoS control in communication networks, RL methods have been applied for single link admission control in ATM networks [10,11] and channel assignment in cellular networks [12]. To the best of our knowledge, the work reported in this paper is the first to use RL for resource management in a DiffServ network.

4

Reinforcement Learning-Based Adaptive Marking (RLAM)

4.1

Motivation

In the proposed RLAM scheme, a novel approach to provide assured end-to-end QoS to flows has been designed, i.e. through adaptive marking of DSCP values in IP packets to select different PHBs and PDBs in different DS domains. An example of how RLAM can be useful would be to consider packets from a session which requires low end-to-end delay, e.g. a Voice over IP session. Typically, packets in this session will be marked with the DSCP value corresponding to the EF or AF PHB. However, in lightly-loaded parts of the network, it may be possible for the BE PHB to satisfy the end-to-end delay requirement. Hence, if packets are marked with the DSCP value corresponding to the BE PHB in those parts of the network, cost savings can be realized since the user or service provider is usually charged a lower rate per bit transmitted for the BE PHB The reinforcement r is the net value of any positive reward that is awarded, e.g. when QoS is satisfied, less any cost or penalty, e.g. the cost of using the PHB over a particular link or DS domain, penalty from QoS violation etc.

Adaptive Marking in Differentiated Services Networks

241

BB2 from other DS domain

Dest D3

R2-IN1

R2-EG1

BB1

BB4

R4-EG1

R2-IN2

Dest D2

Domain 2

Source S1 R1-IN1

R2-EG2

R1-EG1

BB3 R1-EG2

R1-IN2

Source S2

R3-IN1

R4-EG2 R4-IN1

Dest D1 R3-EG1 R4-IN2

Domain 1

Domain 4

Source S3 R3-IN2 from other DS domain

Domain 3 Flow from source to destination Communication between BB and router Reinforcement and Q-value feedback Marking instruction for ingress router Throughput and/or Delay Measurement Probe

R3-EG2 to other DS domain

Core Router Ingress Router Egress Router BB

Bandwidth Broker (BB)

Fig. 1. Data forwarding and feedback paths in the RLAM scheme. compared to the EF PHB. On the other hand, in medium to heavily-loaded parts of the network, the packet may need to be marked as requiring the EF PHB in order to satisfy the low end-to-end delay requirement. 4.2

Overview of Approach

The basic idea behind the proposed RLAM scheme is to (re)mark the DSCP value of packets arriving at each ingress edge router to a DS domain in such a way that the expected cumulative reinforcement achieved over the end-to-end connection is maximized. The cumulative reinforcement can take into account the extent to which the QoS requirements of the flow have been satisfied, less the costs arising from using different PHBs and PDBs in different DS domains and penalties from QoS violation and packet losses. DSCP marking decisions are made by RL agents in each DS domain. From a logical point of view, there is an RL agent RLd,ip at each ingress interface p of every router i which can perform DSCP (re)marking in DS domain d, typically the ingress edge routers. For the rest of this paper, we shall assume that (re)marking does not take place in core or egress edge routers and the forwarding behavior at routers that are traversed by packets in the DS domain are based on the DSCP marking applied at the ingress edge router. The RL agent can either execute at the respective ingress edge router itself, or all the RL agents in the DS domain can be part of the Bandwidth Broker (BB) [13] for that DS domain. To motivate our discussion, a multi-domain DiffServ network is shown in Figure 1 with details of the data forwarding and reinforcement feedback paths in the RLAM scheme for the case with three source-destination pairs. Flows f are grouped into a finite number of flow types f t according to their QoS requirements, e.g. low end-to-end delay flow type for Voice over IP sessions,

242

C.-K. Tham and Y. Liu

a high throughput flow type for bulk FTP sessions, and a combined low endto-end delay and medium throughput flow type for video conferencing sessions. A source node informs the leaf router (the first router to which the source is connected) in the source domain about the end-to-end QoS requirement of a new flow using some signalling protocol such as RSVP. Subsequently, the leaf router assigns an appropriate flow type to that flow. When a packet from the flow arrives at the leaf router, the flow type is tagged onto the packet. This can be done in a variety of ways, e.g. using the two currently unused (CU) bits in the DSCP field in the IP packet header or specific bit patterns in the codepoint space of the DSCP defined for experimental use, i.e. xxxx11 [14], EXP bits in the MPLS shim header or IP options field in the IP header. Other alternatives include using specific fields in UDP, TCP or RTP headers in higher layers of the protocol stack, but these would incur additional packet processing overhead. In our instantiation of this general design which will be described in Section 5, there are three flow types, hence the two CU bits in the DSCP field are sufficient to indicate the flow type that the packet belongs to. At the ingress edge router of the first and subsequent DS domains, the RL agent corresponding to that router selects a DSCP marking for each flow type at the end of every interval of T seconds and sends the marking instruction to the marker at the ingress edge router. Subsequently, the marker marks all incoming packets of a particular flow type during the next interval with the selected DSCP value. The ingress edge router then forwards the packets to the core routers in the DS domain using the underlying routing protocol. At this time, the packets become part of BAs which include packets from other source nodes coming from different ingress edge routers. Following the DiffServ convention that remarking is not done at the core routers and only packet forwarding using the PHB associated with the DSCP marking is carried out, the decision points in each DS domain for the proposed RLAM scheme are at the ingress edge routers only. The RLAM scheme can be readily extended to the case where remarking is done at core routers by implementing additional RL agents corresponding to these routers. When the packets reach the egress edge router, the normal DS traffic conditioning is applied for traffic in BAs that leave the DS domain. At subsequent downstream DS domains, the operations at ingress edge routers, core routers and egress edge routers are performed in the same way described above until the packets reach their destination. To facilitate clear discussion on the quantities involved in different steps of the RLAM approach, we introduce the (d, ip, jq, f t, k) notation which appears as the subscript of quantities like amount of traffic, loss, Q-value, states etc. The notation refers to a quantity in DS domain d which is relevant between ingress interface p of ingress edge router i and egress interface q of egress edge router j for a particular flow type f t at the k th time interval. If a certain element is not relevant, then that variable is omitted from the subscript. For brevity in

Adaptive Marking in Differentiated Services Networks

243

descriptions, we shall refer to the ingress interface p of ingress edge router i in DS domain d as simply ‘ingress dip router’ and the egress interface q of egress edge router j in DS domain d as simply ‘egress djq router’. 4.3

Measurements

A number of measurements are made at ingress edge routers and destinations, using either measurement probes or some built-in functionality at routers and host machines. The measurements at ingress edge routers that are required are: amount of traffic of each flow type arriving at the ingress dip router that is ‘seen’ by RL agent RLd,ip , denoted by td,ip,f t,k 3 . These measurements are used to determine the state xd,ip,k . At a destination node, the measurements that are required are those related to the end-to-end QoS experienced by that flow such as end-to-end delay, throughput and loss [15]. End-to-end delay may be difficult to measure as it requires either a timestamp on each packet and clock synchronization between the source and the measurement probe, or a field within the packet which accumulates the actual delay experienced at each node. These measurements will be compared against the target QoS parameters for that flow type and the appropriate reward will be generated and sent to the BB of the last encountered DS domain. In addition, the BB communicates with all the edge and core routers in the DS domain and can have a domain-wide view of aggregate traffic flows and packet losses between ingress-egress router pairs in the domain. We assume that the BB is able to provide information on the amount of traffic td,ip,jq,f t,k (in unit of bps) and packet losses ld,ip,jq,f t,k . Note that these quantities are for a flow aggregate which comprises a number of individual flows from multiple sources heading towards multiple destinations. 4.4

States and Actions

The state xd,ip,k at a particular RL agent RLd,ip comprises [ttbg ][ttf t1 ][ttf t2 ] . . . [ttf tNf t ], where ttbg is the traffic intensity of background traffic, ttf tn is the traffic intensity of flow type f tn and Nf t is the total number of defined flow types. Note that RL agent RLd,ip is responsible for making marking decisions for each of the Nf t flow types based on the state information. Hence, the RL agent adds the context [f t] to xd,ip,k whenever it accesses Q-values for a specific state and flow type. The action ad,ip,k has several dimensions, one for each flow type f t. The DSCP marking for each f t is selected from the set of DSCP settings corresponding to available PHBs and PDBs such as EF, AF1 and BE. An example of ad,ip,k is [BE,AF1,EF] for flow types 01, 10 and 11 respectively. 3

Note that td,ip,f t,k is a local measurement and is different from the td,ip,jq,f t,k value reported by the the BB to RLd,ip .

244

4.5

C.-K. Tham and Y. Liu

Aggregation of Reinforcement and Q-Value Feedback

In each domain d, per flow rewards are generated when end-to-end QoS parameters are satisfied for destination nodes in the domain; likewise, per flow penalties are generated for end-to-end QoS violations. The rewards and penalties for flows from the ingress dip router to the egress djq router d are aggregated together with the packet transmission costs and penalties for packet losses incurred for flows traversing the same ingress-egress pair in the same domain to produce the reinforcement signal rd,ip,jq,f t,k . In addition, Q-value feedback messages are received from downstream DS domains. All the reinforcement and Q-value feedback messages received by an RL agent RLd,ip are used to update its Q-value. Instead of forwarding all of these messages to upstream DS domains, only the V (y) or Q(y, l) value (see Equations 1 and 2) of that RL agent is fed back to the BB of the previous DS domain for dissemination to the RL agents at the ingress edge routers of that DS domain. Hence, the Q-value passed back by RLd,ip summarizes the ‘goodness’ of subsequent states and DS domains. Traffic that enters the ingress dip router may be split into different proportions to multiple egress djq routers and subsequently to different destinations and downstream DS domains. When reinforcement and Q-value feedback messages from these downstream entities return to DS domain d, their influence on the Q-value update of RL agent RLd,ip are weighted by the equivalent number of flows from ingress dip to egress djq. The prediction error εd,ip,jq,f t,k determined from feedback messages from the egress djq router to RL agent RLd,ip for flow type f t is εd,ip,jq,f t,k = rd,ip,jq,f t,k + γQdjq,ipjq,f t,k (ydjq,ipjq,k , ldjq,ipjq,k ) − Qd,ip,jq,f t,k (xd,ip,k , ad,ip,k )

(4)

where djq and ipjq terms refer to the downstream DS domain and the ingress router in that domain which are connected directly to the egress djq router of the current domain. Finally, the Q-value at the ingress dip router for flow type f t is updated according to: Qd,ip,f t,k+1 (xd,ip,k , ad,ip,k ) = Qd,ip,f t,k (xd,ip,k , ad,ip,k )   td,ip,jq,f t,k ηk εd,ip,jq,f t,k + ARf t q j

(5)

where ARf t is the average rate of flow type f t. In each interval, this procedure is repeated for the other flow types at the ingress dip router followed by the RL agents at other ingress ports in other ingress edge routers in domain d.

Adaptive Marking in Differentiated Services Networks

5 5.1

245

ns2 Implementation Network Topology

The network shown in Figure 2 together with the RLAM scheme described above have been implemented using ns2 [16].

Dest N1 Dest N2

Source N3

Dest N3 Source N2

BB1

BB2

BB3

Source N1

Source F1

Source F2 Source F3

Dest F1

Dest F2

Dest F3

Domain 1

Domain 2

Domain 3

Throughput and/or Delay Measurement Probe Feedback of reinforcement or Q-value to previous RL agent Marking instruction for ingress edge router

BB

Core Router Ingress Edge Router Egress Edge Router Bandwidth Broker

Fig. 2. ns2 implementation of network with three DS domains. There are three flow types (Nf t = 3) with end-to-end QoS specifications: 1. High throughput required (≥ 128 Kbps) 2. Low delay required (< 100 ms end-to-end) 3. Moderate throughput (≥ 64 Kbps) and low delay required (< 200 ms end-toend) Note that these flow types represent the types of assured service offered by the network. These flow types are indicated in each packet using the CU bits in the DSCP field with values 01, 10 and 11 respectively, i.e. f t ∈ {01, 10, 11}. Packets from flows corresponding to background traffic which are not handled by the RLAM scheme will have the value 00 in their CU bits. 5.2

Traffic Characteristics

Traffic is generated by sources which represent user applications and their characteristics are shown in Table 1. We consider three types of traffic: (1) Bulk FTP with the average rate shown in the table and no delay requirement (sent using TCP); (2) Voice over IP (VoIP) sessions which are modelled as ON-OFF sources with the peak rate as shown and an end-to-end delay requirement of less than 150 ms (sent using UDP); and (3) video conferencing sessions also modelled as ON-OFF sources with a higher peak rate and end-to-end delay requirement of less than 200 ms (sent using UDP). VoIP and video conferencing traffic are ON/OFF sources with the same ON time (500 ms) and OFF time (500 ms); the holding time for all flows is 30 seconds. The appropriate flow types for these sessions would be 01, 10 and 11, respectively, which will be tagged onto packets by the leaf or edge router. In addition, different types of background traffic are generated and tagged with CU bits 00.

246

C.-K. Tham and Y. Liu

Table 1. Characteristics of the three traffic types. Traffic Type Bulk FTP VoIP Video Conf TCP bckgrd VoIP bckgrd Video bckgrd

6

CU Arrival Peak Average Packet bit Rate Rate Rate Size (s−1 ) (Kbps) (Kbps) (bytes) 01 1/10 / 128 1,000 10 1/3 64 32 150 11 1/15 128 64 150 00 1/15 / 128 1,000 00 1/15 64 32 150 00 1/15 128 64 150

Simulation Results

Simulations using the ns2 implementation described above have been carried out with the same traffic conditions for three marking schemes: (1) the proposed Reinforcement Learning Adaptive Marking (RLAM) scheme, (2) Static Marking (SM), in which all packets with flow types 01, 10 and 11 will be marked statically with the DSCP value corresponding to BE, AF1 and EF respectively, and (3) Random Marking (RM), in which the marking for these flow types will be selected randomly from the three DSCP values at the ingress router of each DS domain. Each simulation lasts for 6,000 seconds. 15000

RLAM Static Marking Random Marking

300000

10000

250000

5000

Total reinforcement received

Total packet transmission cost incurred

350000

200000

150000

100000

50000

0

RLAM

0

-5000

-10000

-15000

0

1000

2000

3000 Time(seconds)

4000

5000

6000

(a) Total cumulative packet transmission cost for the RLAM, SM and RM schemes.

-20000

0

1000

2000

3000 Time(seconds)

4000

5000

6000

(b) Total cumulative reinforcement received in RLAM scheme.

Fig. 3. Performance of the RLAM scheme. Table 2. QoS achieved and transmission cost incurred using RLAM scheme Traffic Type

Average Average Throughput Delay (bps) (ms) Bulk FTP 125,935 59.9 VoIP 30,819 47.1 Video Conf 61,433 58.7

Total Transm. Loss Cost (pkts) Incurred 39 8,137 137 4,873 103 2,637

Adaptive Marking in Differentiated Services Networks

247

Table 3. QoS achieved and transmission cost incurred using SM scheme Traffic Type

Average Average Throughput Delay (bps) (ms) Bulk FTP 125,739 58.9 VoIP 30,679 45.7 Video Conf 63,053 57.2

Total Transm. Loss Cost (pkts) Incurred 0 5,195 1 14,363 1,131 17,527

Table 4. QoS achieved and transmission cost incurred using RM scheme Traffic Type

Average Average Total Transm. Throughput Delay Loss Cost (bps) (ms) (pkts) Incurred Bulk FTP 125,665 58.8 4 26,054 VoIP 30,779 47.4 120 16,990 Video Conf 60,871 57.6 165 8,801

The simulation results for the three schemes will be presented in the following format. First, the total packet transmission cost for each of the three schemes will be presented. This will be followed by a discussion of the behavior of the proposed RLAM scheme. Lastly, the QoS achieved and the packet transmission cost over a defined period for each of the three schemes will be examined. The total cumulative packet transmission cost for the six traffic sources of interest excluding the background traffic for each of the three schemes over the 6,000 seconds of simulation time can be seen in Figure 3(a). Throughout the whole period, it can be seen that the RM marking scheme incurs the highest cost. This is because a large number of packets from the bulk FTP sessions have used the higher PHBs such as EF and AF1 even when it is not necessary to do so in order to satisfy their QoS requirements. In the early stages when t < 400 s, the cost incurred by RLAM is the same as that for the RM scheme, showing that marking action selection in RLAM is random at that time. When t < 1,400 s, the cost incurred by the RLAM scheme is slightly higher than that incurred by the SM scheme as the RL agents are still in their exploration and training phase. After 1,400 seconds, the total transmission cost for RLAM becomes lower than that for the SM and RM schemes, showing that RLAM has learnt to select cost-efficient PHBs. As time goes on, the difference in total transmission cost between the three schemes continues to increase, with RLAM incurring significantly lower cost compared to the other two schemes. Next, we examine the variation in the total cumulative reinforcement received by the three RL agents in the RLAM scheme over the 6,000 second simulation period (Figure 3(b)). In the first 1,400 seconds, the RL agents encounter high costs and penalties due to QoS violations and selection of expensive PHBs and the net reinforcement received per unit time is negative. Between 2,000 to 2,700 seconds, the rewards received per unit time balance the costs and penalties

248

C.-K. Tham and Y. Liu

incurred per unit time, hence the total reinforcement curve is flat during this period. After that, the total reinforcement curve increases almost linearly since the net reinforcement received per unit time is positive most of the time. This indicates that the QoS associated with each flow type are satisfied for most of the flows, i.e. the RL agents are selecting the appropriate PHBs for each flow type in each DS domain in order to provide the desired end-to-end QoS. Since the objective of the RLAM scheme is to satisfy the QoS requirements associated with the flow types in a cost-effective way, we compare the QoS achieved and packet transmission cost incurred for the different traffic types from t = 5,000 to 6,000 s when the traffic conditions are stable and the RLAM scheme has converged. Tables 2, 3 and 4 show the average throughput and average delay per flow, and the total packet loss and transmission cost incurred for each of the three different traffic types. As expected, the video conferencing sessions achieved significantly higher throughput when using the SM scheme compared to the other 2 schemes. This is due to the EF PHB, although some packet losses occurred since out-of-profile packets are discarded in the EF PHB. Other than that, since the network is moderately loaded, it can be seen that the average throughput and average delay of the corresponding traffic type under the three marking schemes are similar. Most significantly, the total packet transmission cost incurred in this interval for the RLAM scheme is less than half that of the SM scheme and less than onethird that of the RM scheme, with most of the savings coming from being able to find a more cost effective way to carry VoIP and video conferencing traffic without severely violating the end-to-end QoS requirements. The utilization of each of the provisions for the BE BA, AF1 BA and EF BA respectively, which includes the background traffic, in one of the links in Domain 2 for the three marking schemes are: (1) RLAM: 49.36%, 5.87%, 9.60% (2) SM: 26.74%, 17.51%, 16.53% (3) RM: 22.86%, 19.00%. 18.38%. Thus, the RLAM scheme has used more of the low cost BE PHB compared to the other two marking schemes to carry the traffic, thus enabling it to achieve significant cost savings. Note that RLAM has reached a balance and does not attempt to send all the traffic using the BE PHB since that would lead to penalties arising from QoS violations and packet losses.

7

Conclusion

In this paper, a Reinforcement Learning-based Adaptive Marking (RLAM) scheme has been proposed and its design and implementation considerations explained. Simulations done using ns2 show that the RLAM scheme is effective in providing end-to-end QoS to different user applications such as VoIP and video conferencing at a significantly lower total packet transmission cost compared to the commonly used static marking approach. In future work, we plan to improve the speed of convergence of the RLAM algorithm through the use of the TD(λ) temporal differences [17] algorithm as well as investigate the effectiveness of different ways of representing state information [18].

Adaptive Marking in Differentiated Services Networks

249

References 1. S. Shenker, C. Partridge and R. Guerin, Specification of Guaranteed Quality of Service, IETF RFC 2212, Sept 1997. 237 2. S. Blake, et al, An Architecture for Differentiated Services, IETF RFC 2475, Dec 1998. 237 3. V. Jacobson, et al, An Expedited Forwarding PHB, IETF RFC 2598, June 1999. 237 4. J. Heinanen, et al, Assured Forwarding PHB Group, IETF RFC 2597, June 1999. 237 5. F. Wang, P. Mohapatra and D. Bushmitch, A Random Early Demotion and Promotion Marker for Assured Services, IEEE Jour. on Selected Areas in Communications, vol. 18, no. 12, Dec 2000. 239 6. W. C. Feng, D. D. Kandlur, D. Saha and K. G. Shin, Adaptive Packet Marking for Maintaining End-to-End Throughput in a Differentiated-Services Internet, IEEE/ACM Trans. on Networking, vol. 7, no. 5, Oct 1999. 239 7. A. Barto, R. Sutton and C. Anderson, Neuron-like Elements That Can Solve Difficult Learning Control Problems, IEEE Trans. on Systems, Man and Cybernetics, vol. 13, pp. 835-846, 1983. 239 8. D. P. Bertsekas and J. N. Tsitsiklis, Neuro-Dynamic Programming, Athena Scientific, Belmont, MA, USA, 1996. 239 9. C. J. C. H. Watkins and P. Dayan, Q-Learning, Machine Learning, vol. 8, pp. 279-292, 1992. 239 10. H. Tong and T. X. Brown, Adaptive Call Admission Control Under Quality of Service Constraints: A Reinforcement Learning Solution, IEEE Jour. on Selected Areas in Communications, vol. 18, no. 2, Feb 2000. 240 11. P. Marbach, O. Mihatsch and J. N. Tsitsiklis, Call Admission Control and Routing in Integrated Services Networks using Neuro-Dynamic Programming, IEEE Jour. on Selected Areas in Communications, vol. 18, no. 2, Feb 2000. 240 12. J. Nie and S. Haykin, A Dynamic Channel Assignment Policy Through Q-Learning, IEEE Trans. on Neural Networks, vol. 10, no. 6, Nov 1999. 240 13. F. Reichmeyer, L. Ong, A. Terzis, L. Zhang and R. Yavatkar, A Two-Tier Resource Management Model for Differentiated Services Networks, IETF Internet Draft 2tier-draft, Nov 1998. 241 14. K. Nichols, S. Blake, F. Baker and D. Black, Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers, IETF RFC 2474, Dec 1998. 242 15. W. Jiang and H. Schulzrinne, QoS Measurement of Internet Real-Time Multimedia Services, Technical Report CUCS-015-99, Dept of Comp. Sc., Columbia University, 1999. 243 16. S. McCanne and S.Floyd, ns2 - The Network Simulator, available from http://www.isi.edu/nsnam/ns/. 245 17. R. S. Sutton, Learning To Predict by The Methods of Temporal Differences, Machine Learning, vol. 3, pp. 835-846, 1988. 249 18. C. K. Tham, Reinforcement Learning of Multiple Tasks using a Hierarchical CMAC Architecture, Robotics and Autonomous Systems, Special Issue on Reinforcement Learning and Robotics, vol. 15, pp. 247-274, Elsevier, July 1995. 249

Dynamic QoS Adaptation Using COPS and Network Monitoring Feedback Toufik Ahmed1,2, Ahmed Mehaoua1, and Raouf Boutaba2 1

2

University of Versailles, CNRS-PRiSM Lab. 45 av. des Etats-Unis, 78000, Versailles, France {tad, mea}@prism.uvsq.fr University of Waterloo, Dept. of Computer Science 200 University Avenue West, Waterloo,

Ont. N2L 3G1, Canada {tad, rboutaba}@bbcr.uwaterloo.ca

Abstract. This paper presents an approach to handle out of profile traffic using Common Open Policy Service and network monitoring feedback. The proposed approach is based on monitoring and reporting information sent by bandwidth monitors installed on each node of a Diffserv Domain. A monitor interacts with Policy Decision Point. This later, depending on the network state, pushes policy decision rules to the Policy Enforcement Point in order to accept, remark or drop out-of-profile traffic dynamically. This allows a dynamic reallocation and management of network resources based on current network state and applications QoS requirements. An implementation and a performance evaluation of the dynamic QoS adaptation framework using a Java COPS and a Linux-based network testbed are also presented. Keywords: IP Diffserv, QoS Adaptation, COPS, Network Monitoring.

1

Introduction

Recent works on IP Quality of Service (QoS) Management led to the development and standardization of enhanced protocols and services. The IETF has defined the Policy-based Network Management (PBNM) architecture to configure network services. Currently most efforts are focused on Differentiated Services (Diffserv) in the Internet. The goal of the policy-based network management is to enable network control and management on a high abstraction level by defining configuration rules called policies. Policies specify how a network node must be configured in vendorindependent, interoperable and scalable manner. Diffserv architecture defines, at a lower level, four types of data-path elements: traffic classifiers, actions elements, meters and queuing elements [1]. Combining these elements into higher-level blocks creates a Traffic Condition Block (TCB), which can be managed by policy-based network management tools. The configuration of Diffserv TCB using PBNM involves the use of administratively prescribed rules that specify actions in response to defined criteria. All the information needed to perform this task such as profiles, user information, network configuration data, and IP infrastructure data such as network addresses and name server information are The original version of this chapter was revised: The copyright line was incorrect. This has been corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-540-45812-8_28 K.C. Almeroth and M. Hasan (Eds.): MMNS 2002, LNCS 2496, pp. 250-262, 2002. © IFIP International Federation for Information Processing 2002

Dynamic QoS Adaptation Using COPS and Network Monitoring Feedback

251

stored in a policy repository. These configurations do not change frequently because they are not associated with specific application or traffic but with the network management. The more difficult part in the configuration is to have the traffic entering the network appropriately marked (audio, video and other data). Since, the user is signed up for the service, edge devices could be configured to mark userís traffic with the appropriate PHB. With a known IP address and/or IP port number, the administrator can specify a policy that refers to user application IP address and marks traffic coming from that address appropriately. In the actual configuration, when the user signs up for a particular service, he/she must specify his/her traffic profile and the action that must be taken when the traffic exceed this predefined profile (out of profile traffic). Generally, the out of profile traffic is dropped or marked as best effort traffic. This model is static and does neither respond to application needs nor favor an optimal utilization of network resources. In this paper a Diffserv QoS management is explored with the objective to overcome the limitations of the static model. In addition to the techniques described above (i.e., PBNM, TCB, etc.), network monitors are used to make the system reactive by making automatic and real-time decisions concerning out of profile traffic. An architectural model allowing the configuration and dynamic management of the Diffserv domain will be presented, experimented and evaluated. The reminder of this paper is as follows. In section 2, we compare static and dynamic policy decision approaches and present our proposal, which is based on dynamic QoS adaptation through network resource monitoring and feedback signaling. Section 3 is devoted to the implementation of the proposed QoS network management framework. Performance evaluation and results analysis are discussed in Section 4. Finally, we conclude in Section 5.

2

Dynamic QoS Adaptation

2.1

Static Policy Decision

In the Diffserv architecture, a particular traffic receives a predefined treatment based on predefined policies. This treatment is interpreted as a particular PHB [2], [3]. This task is done by the TC (Traffic Control) function, which assigns the correct DSCP [4] for the clientís traffic according to it SLA (Service Level Agreement). Recall that each client defines it requirements and these are translated into SLAs. The allocation of resources (QoS) still static and can lead to bandwidth wasting and starving clients. Some algorithms such as Time Sliding Window Three Colour Marker (TSWTCM) [5] and a Two Rate Three Color Marker (TRTCM) [6] can be used to mark IP packets treated by the edge router with a Diffserv PHB. These algorithms meter the traffic stream and marks packets based on measured throughput. To receive a particular treatment, the user must specify it profile TSpec (Traffic Specification). TSpec specifies the temporal properties of a traffic stream selected by a classifier. It provides rules for determining whether a particular packet is in profile or out of profile. The Meter uses a Token Bucket to control user traffic. The following is a non-exhaustive list of potential profile parameters: 1. Peak rate p in bits per sec (bps) 2. Token bucket rate r (bps), 3. Bucket depth b (bytes),

252

T. Ahmed, A. Mehaoua, and R. Boutaba

An Excess Treatment parameter describes how the service provider will process excess traffic, i.e. out of profile traffic. The process takes place after Traffic Conformance Testing. Excess traffic may be dropped, shaped and/or remarked. Depending on the particular treatment, more parameters may be required, e.g. the DSCP value in case of re-marking or the shapers buffer size for shaping. All these actions are decided once the network element is configured and are not changed over the time. Fig. 1 gives an example of how out of profile traffic is treated using static configuration. In this Figure, user sends traffic not conforming to his Traffic Specification. Edge router control this traffic by a token bucket. Non-conforming traffic will be dropped always.

f ix e d b a d w id th

Drop always

Input traffic

Edge router control

Output traffic

Fig. 1. Static Policy Decision For this reason, there is a great need to control dynamically the action taken by the network element for more flexible resource allocation. For this, different conditioning actions may be performed on the in profile packets and out of profile packets or different accounting actions may be triggered dynamically according to current network state. Clearly, a more flexible resource allocation can be achieved by controlling dynamically network elements behavior. 2.2

Dynamic Policy Decision

In the static approach out of profile traffic is simply dropped, remarked or assigned a new profile. This decision is static and is taken once for all, i.e. when the network element is configured. For example the Policing Rule = drop out-of profile packets can be applied to all the packets which are out of profile regardless of whether the network is capable or not to transmit this packet. Fig. 2 shows where we can dynamically decide what actions must be applied to out of profile packets. In contrast to the static approach, these actions vary according to the network state (network link load, traffic behavior, etc.).

Dynamic QoS Adaptation Using COPS and Network Monitoring Feedback

2.3

253

Automatic Diffserv Domain Configuration

When a network element is started, its local PEP requests the PDP for all policies concerning Diffserv traffic marking using COPS (Common Open Policy Service) [7], [8], [9]. The policies sent by the PDP to the PEP, may concern entire router QoS configuration or a portion of it, as an updating of a Diffserv marking filter. The PDP may proactively provision the PEP reacting to external events generated by some monitors such as a bandwidth monitor. r bps

User Profile (Token Bucket) b bits

Traffic

In Profile

Meter

Dropping

Dynamic Decision

Accepting traffic Re-Marking (BE, AF, EF, etc)

Out-Of-Profile

Fig. 2. Dynamic Decision

PDP

New policy decisions

3

LDAP

1: Installing filter 2: Monitoring event 3: Provisionning operation

1

out of rofile traffic

Policy decision

2

Event

PEP Monitoring agent Edge Router

QoS Monitor

Edge Router

Monitoring agent PEP: Policy Enforcement Point PDP: Policy Decision Point

Fig. 3. COPS-PR with Monitoring Event Fig. 3 shows the steps involved in the configuration of Diffserv domain. These steps are as follow: − Step 1: When the edge router is started, the local PEP requests all policy decisions concerning Diffserv QoS Management (filtering, classes, queuing discipline, and actions for out of profile traffic). All incoming traffics are processed according to the pre-installed rules.

254

T. Ahmed, A. Mehaoua, and R. Boutaba

− Step 2: When the bandwidth monitor, installed on the core router, detects a significant change in the amount of available bandwidth, it triggers an external event reported to the PDP indicating the current bandwidth availability. − Step 3: The PDP pushes to the edge router (PEP) an update of QoS Management decision. These steps allow configuring correctly different policies related to the same traffic. We introduce the following policy rule: let us On event: If then . A Profile is used to determine when a policy rule applies, for instance given pairs of source and destination addresses. An Action is a performed by that the policy enforcement entity to traffic of a given profile. Examples of actions are marking, accepting or rejecting traffic. Example of policy rules: • Rule 1: Mark DSCP value EF on all packets with source addresses from 193.51.25.1 to 193.51.25.255 priority 0 • Rule 2: Mark DSCP value AF11 on all packets with destination address 200.200.200.100 priority 1 2.4

Example of Application

Assume, an audio application has subscribed to a particular Diffserv class (an Expedited Forwarding Class). Audio traffic is defined by a particular profile. In this example Diffserv class simply mean that the audio stream will be marked with the appropriate DSCP (EF PHB here). The Administrator of the Diffserv domain configures the environment to support the Gold, Silver, Bronze and other services. Such configuration can be done through a Bandwidth Broker. Supporting different classes of service in the core network requires putting in place classifiers, which cause the devices to examine the Diffserv mark on the packet and then treat the traffic accordingly. These configurations do not change frequently because they are not associated with specific application or traffic but with the network management. Since, the application is signed up for the service, edge devices are configured to mark applicationís traffic with the appropriate PHB. Based on the IP address and/or the port number, the administrator can set a policy that marks traffic coming from that address with EF PHB. In order for customized traffic going to audio application (e.g. feedback traffic, RTCP, client commands) to receive a Diffserv treatment, policy must be deployed to the opposite edge device of a Diffserv domain. When the audio application starts sending the data, the edge router must ensure; (1) the data sent by the audio server does not exceed what the application has subscribeto (SLA) and (2) marking conforming traffic (in profile traffic) with the appropriate PHB (EF PHB in our example). In case of receiving out of profile traffic, the edge router requests a decision from the PDP. Since the PDP knows the current network state - because it receives monitoring information from different monitors installed in the network, it decides a new policy rule, for example dropping, marking or accepting out of profile traffic. This decision varies according to current network state.

Dynamic QoS Adaptation Using COPS and Network Monitoring Feedback

2.5

255

QoS Management Algorithm

We have configured 3 rules named Rule1, Rule2 and Rule3 to deal with out of profile traffic. The Policy Server can choose one rule among the several depending on information sent periodically by the monitors. The monitoring information concerns essentially the bandwidth usage of each link in the network. The calculation of shaped value of the bandwidth using Exponentially Weighted Moving Average (EWMA) is presented in section 2.6. Below in Fig. 4 our algorithm, which uses the predefined policy rules to make a decision depending on bandwidth usage in the network. Initialization: Start Bandwidth Monitor Mi for each Router i to calculate the available bandwidth BWi Lambda ← 0.2 // Fixed value for historical data X ←50% link capacity // Initial value of EWMA Min_th← 40% link capacity Max_th← 70% link capacity Loop: BW ← max(BW1, BW2, …,BWi) //EWMA available bandwidth X X← (1-lambda) * BW + lambda * X if X < Min_th then Rule1: Accept out-of-profile traffic else if Min_th